# Ken Imoto

> AI Systems Engineer specializing in LLMO, WebRTC, Real-time AI, and Context Engineering. Building AI-native organizations powered by LLMs, automation, and distributed agents.

## About

Ken Imoto is an AI Systems Engineer based in Fukuoka, Japan. He builds AI-native organizations at Propel-Lab (https://propel-lab.co.jp) and researches LLMO (LLM Optimization), AI Agent Design, and Human-AI Interaction.

LLMO includes approaches such as AEO (Answer Engine Optimization) and GEO (Generative Engine Optimization).

## Expertise

- LLMO (LLM Optimization) — Author of the LLMO Framework (https://llmoframework.com)
- WebRTC & Real-time AI
- Context Engineering (CLAUDE.md, AGENTS.md patterns)
- AI Agent Design (MCP, multi-agent architectures)
- Android & Web Development (8+ years)

## Books (36 Kindle editions, 11 Zenn Books, auto-generated)

### English

- [Chaos Engineering in Practice — From Netflix's monkey to AI-driven experiments — Landing Page](https://kenimoto.dev/books/chaos-engineering-guide/)
- [Debugging Engineering — Ten habits, translated into prompts, layered with CLAUDE.md and hooks — Landing Page](https://kenimoto.dev/books/debugging-engineering/)
- [Cognitive Bias for Software Engineers — 20 Biases That Distort Estimates, Code Reviews, and Team Decisions — Landing Page](https://kenimoto.dev/books/engineer-psychology-tricks/)
- [Cognitive Bias for Software Engineers — Kindle](https://www.amazon.com/dp/B0HH99235S)
- [Why Do Some Words Stay With You Forever? — 100 Engineering Quotes Decoded — Landing Page](https://kenimoto.dev/books/engineer-it-quotes/)
- [Why Do Some Words Stay With You Forever? — Kindle](https://www.amazon.com/dp/B0GX3CS95M)
- [The Practical Knowledge Graph Guide — Structure Your Data, Sharpen Your AI — Landing Page](https://kenimoto.dev/books/knowledge-graph-practical-guide/)
- [The Practical Knowledge Graph Guide — Kindle](https://www.amazon.com/dp/B0GX2Z73JV)
- [The 300ms Threshold — Why Talking to AI Feels Wrong — Landing Page](https://kenimoto.dev/books/voice-ai-300ms-ux/)
- [The 300ms Threshold — Kindle](https://www.amazon.com/dp/B0GYQ4L2KP)
- [Systematizing AI Code Review — The 3-Layer Model for 60% Faster Reviews — Landing Page](https://kenimoto.dev/books/harness-code-review/)
- [Systematizing AI Code Review — Kindle](https://www.amazon.com/Systematizing-AI-Code-Review-Propel-Lab-ebook/dp/B0GX71K2BV)
- [LLMO Quickstart — AI Search Optimization for Engineers — Landing Page](https://kenimoto.dev/books/llmo-quickstart/)
- [LLMO Quickstart — Kindle](https://www.amazon.com/dp/B0GHJLJ611)
- [MCP Security in Practice — What OWASP Won't Tell You About AI Tool Integrations — Landing Page](https://kenimoto.dev/books/mcp-security-practice/)
- [Harness Engineering — From Using AI to Controlling AI — Landing Page](https://kenimoto.dev/books/harness-engineering-guide/)
- [Harness Engineering — Kindle](https://www.amazon.com/dp/B0FZNL8D1V)
- [Why ChatGPT Ignores Your Website — The LLMO Practical Guide — Landing Page](https://kenimoto.dev/books/llmo-ai-search-optimization/)
- [Why ChatGPT Ignores Your Website — Kindle](https://www.amazon.com/dp/B0GQ2L5HQY)
- [Turning LLMs from Liars into Experts — Context Engineering in Practice — Landing Page](https://kenimoto.dev/books/context-engineering/)
- [Turning LLMs from Liars into Experts — Kindle](https://www.amazon.com/dp/B0GHRJX36Z)
- [Practical Claude Code — Context Engineering for Modern Development — Landing Page](https://kenimoto.dev/books/claude-code-mastery/)
- [Practical Claude Code — Kindle](https://www.amazon.com/dp/B0GPWPKNB7)

### Japanese (日本語)

- [カオスエンジニアリング実践ガイド — Netflixの猿からAI駆動の実験へ — Landing Page](https://kenimoto.dev/ja/books/chaos-engineering-guide/)
- [Claude Code で動かすマイコン — AI エンジニアが組み込みに踏み込む実装入門 — Landing Page](https://kenimoto.dev/ja/books/claude-code-hardware-control/)
- [デバッグの技法 — 10の思考武器と現場の物語 — Landing Page](https://kenimoto.dev/ja/books/debugging-engineering/)
- [エンジニアのための多層防御設計 — 穴のないチーズは存在しない — Landing Page](https://kenimoto.dev/ja/books/multi-layer-defense-design/)
- [トレースとログの教科書 — フロントエンドとバックエンドで何が違うのか — Landing Page](https://kenimoto.dev/ja/books/observability-frontend-backend/)
- [人流分析実践ガイド — Pythonで読み解く都市の動き — Landing Page](https://kenimoto.dev/ja/books/people-flow-analysis-guide/)
- [なぜNVIDIAはロボットに全賭けしたのか — エンジニアのためのフィジカルAI入門 — Landing Page](https://kenimoto.dev/ja/books/physical-ai-intro/)
- [リファクタリングの教科書 — React / TypeScript / Python 実践パターン — Landing Page](https://kenimoto.dev/ja/books/refactoring-patterns/)
- [声が届く範囲を設計する — Landing Page](https://kenimoto.dev/ja/books/speech-privacy-design/)
- [ビデオ通話アプリの自動テスト — Playwrightで2台分のテスターを置き換える — Landing Page](https://kenimoto.dev/ja/books/webrtc-e2e-playwright/)
- [AIくさい文章から脱出する技術 — 6モデル180サンプルの実測でわかった「バレる文章」の正体と差別化の手順 — Landing Page](https://kenimoto.dev/ja/books/ai-text-slop-escape/)
- [AIくさい文章から脱出する技術 — Kindle](https://www.amazon.co.jp/dp/B0H75RMJXT)
- [8000行でわかる大規模言語モデル — Karpathyのnanochatで追う、事前学習から推論エンジンまでの全工程 — Landing Page](https://kenimoto.dev/ja/books/nanochat-code-reading/)
- [8000行でわかる大規模言語モデル — Kindle](https://www.amazon.co.jp/dp/B0HCH7VTBX)
- [8000行でわかる大規模言語モデル — Zenn](https://zenn.dev/kenimo49/books/nanochat-code-reading)
- [RTX 4070でQwen 35Bを2.8倍速くする — ローカルLLM 高速化・チューニングの実測ノート llama.cpp 実践ガイド — Landing Page](https://kenimoto.dev/ja/books/local-llm-qwen-4070/)
- [RTX 4070でQwen 35Bを2.8倍速くする — Kindle](https://www.amazon.co.jp/dp/B0HCH7MGW1)
- [RTX 4070でQwen 35Bを2.8倍速くする — Zenn](https://zenn.dev/kenimo49/books/local-llm-qwen-4070)
- [AIに選ばれる店をつくる — 店舗オーナーのための AI Native MEO 実践 — Landing Page](https://kenimoto.dev/ja/books/store-owner-ai-meo-llmo/)
- [AIに選ばれる店をつくる — Kindle](https://www.amazon.co.jp/dp/B0H3VX2YZC)
- [エンジニアの心理トリック大全 — 認知バイアスから読み解くエンジニアの実務心理学 — Landing Page](https://kenimoto.dev/ja/books/engineer-psychology-tricks/)
- [エンジニアの心理トリック大全 — Kindle](https://www.amazon.co.jp/dp/B0HCJB8MF3)
- [エンジニアの心理トリック大全 — Zenn](https://zenn.dev/kenimo49/books/engineer-psychology-tricks)
- [エンジニアリング100の言葉 — なぜその一文は記憶に残るのか — Landing Page](https://kenimoto.dev/ja/books/engineer-it-quotes/)
- [エンジニアリング100の言葉 — Kindle](https://www.amazon.co.jp/dp/B0GXR5Q6NW)
- [ナレッジグラフ活用大全 — 構造化すれば、AIは賢くなる — Landing Page](https://kenimoto.dev/ja/books/knowledge-graph-practical-guide/)
- [ナレッジグラフ活用大全 — Kindle](https://www.amazon.co.jp/dp/B0GX465PG7)
- [音声AIの300ms — 人はなぜAIとの会話に違和感を覚えるのか — Landing Page](https://kenimoto.dev/ja/books/voice-ai-300ms-ux/)
- [音声AIの300ms — Kindle](https://www.amazon.co.jp/dp/B0GX2R5HP4)
- [音声AIの300ms — Zenn](https://zenn.dev/kenimo49/books/voice-ai-300ms-ux)
- [LLMOクイックスタート — エンジニアのためのAI検索最適化入門 — Landing Page](https://kenimoto.dev/ja/books/llmo-quickstart/)
- [LLMOクイックスタート — Kindle](https://www.amazon.co.jp/dp/B0GHMCKQPX)
- [AIコードレビューを仕組み化する技術 — hooks・AI・人間の3層モデル — Landing Page](https://kenimoto.dev/ja/books/harness-code-review/)
- [AIコードレビューを仕組み化する技術 — Kindle](https://www.amazon.co.jp/dp/B0GHT7FQ7G)
- [AIコードレビューを仕組み化する技術 — Zenn](https://zenn.dev/kenimo49/books/harness-code-review)
- [MCP実践セキュリティ — 本番導入で躓かないための完全ガイド — Landing Page](https://kenimoto.dev/ja/books/mcp-security-practice/)
- [なぜAI生成UIは全部青いのか — 同質化から脱出する技術 — Landing Page](https://kenimoto.dev/ja/books/ai-slop-escape-guide/)
- [なぜAI生成UIは全部青いのか — Kindle](https://www.amazon.co.jp/dp/B0GNCNLK9K)
- [なぜAI生成UIは全部青いのか — Zenn](https://zenn.dev/kenimo49/books/ai-slop-escape-guide)
- [LLMOに最適化されたホームページをゼロから作る — GitHub Pagesで無料公開、AIに見つけてもらうサイト設計 — Landing Page](https://kenimoto.dev/ja/books/llmo-website-builder/)
- [LLMOに最適化されたホームページをゼロから作る — Zenn](https://zenn.dev/kenimo49/books/llmo-website-builder)
- [ハーネス・エンジニアリング — AIを\"使う\"から\"操る\"へ — Landing Page](https://kenimoto.dev/ja/books/harness-engineering-guide/)
- [ハーネス・エンジニアリング — Kindle](https://www.amazon.co.jp/dp/B0GF8VGSPC)
- [Claude Codeクイックスタート — 30分で始めるAIエージェント開発 — Landing Page](https://kenimoto.dev/ja/books/claude-code-quickstart/)
- [Claude Codeクイックスタート — Zenn](https://zenn.dev/kenimo49/books/claude-code-quickstart)
- [なぜあなたのサイトはChatGPTに無視されるのか — LLMO実践ガイド — Landing Page](https://kenimoto.dev/ja/books/llmo-ai-search-optimization/)
- [なぜあなたのサイトはChatGPTに無視されるのか — Kindle](https://www.amazon.co.jp/dp/B0GQ4XPBMJ)
- [なぜあなたのサイトはChatGPTに無視されるのか — Zenn](https://zenn.dev/kenimo49/books/llmo-ai-search-optimization)
- [LLMを「嘘つき」から「専門家」に変える技術 — Context Engineering 実践入門 — Landing Page](https://kenimoto.dev/ja/books/context-engineering/)
- [LLMを「嘘つき」から「専門家」に変える技術 — Kindle](https://www.amazon.co.jp/dp/B0GHNC29MF)
- [LLMを「嘘つき」から「専門家」に変える技術 — Zenn](https://zenn.dev/kenimo49/books/context-engineering)
- [実践Claude Code — コンテキストエンジニアリングで開発が変わる — Landing Page](https://kenimoto.dev/ja/books/claude-code-mastery/)
- [実践Claude Code — Kindle](https://www.amazon.co.jp/dp/B0GPSVW3BJ)
- [実践Claude Code — Zenn](https://zenn.dev/kenimo49/books/claude-code-mastery)

### Portuguese (Português)

- [Chaos Engineering na Prática — Do macaco da Netflix aos experimentos guiados por IA — Landing Page](https://kenimoto.dev/pt/books/chaos-engineering-guide/)
- [Manual completo de Knowledge Graph — Estruture os dados, torne a IA mais inteligente — Landing Page](https://kenimoto.dev/pt/books/knowledge-graph-practical-guide/)
- [Manual completo de Knowledge Graph — Kindle](https://www.amazon.com.br/dp/B0H3FVPXZV)
- [Por Que Algumas Palavras Ficam para Sempre? — 100 Frases da Engenharia Decifradas — Landing Page](https://kenimoto.dev/pt/books/engineer-it-quotes/)
- [Por Que Algumas Palavras Ficam para Sempre? — Kindle](https://www.amazon.com.br/dp/B0H2KR8L6T)
- [LLMO Quickstart — Otimização para Busca por IA para Engenheiros — Landing Page](https://kenimoto.dev/pt/books/llmo-quickstart/)
- [LLMO Quickstart — Kindle](https://www.amazon.com.br/dp/B0H1YG3QDP)
- [Manual completo dos truques psicológicos para engenheiros — Vieses cognitivos e técnicas psicológicas traduzidos para a prática do engenheiro — Landing Page](https://kenimoto.dev/pt/books/engineer-psychology-tricks/)
- [Manual completo dos truques psicológicos para engenheiros — Kindle](https://www.amazon.com.br/dp/B0GZFBC7SP)
- [Revisão de Código com Harness Engineering — Automação de revisão na era dos agentes de IA — Landing Page](https://kenimoto.dev/pt/books/harness-code-review/)
- [Revisão de Código com Harness Engineering — Kindle](https://www.amazon.com.br/dp/B0H2DB9YXD)
- [Transformando LLMs de Mentirosos em Especialistas — Engenharia de Contexto na Prática — Landing Page](https://kenimoto.dev/pt/books/context-engineering/)
- [Transformando LLMs de Mentirosos em Especialistas — Kindle](https://www.amazon.com.br/dp/B0H12378CJ)
- [Harness Engineering — De Usar IA a Controlar IA — Landing Page](https://kenimoto.dev/pt/books/harness-engineering-guide/)
- [Harness Engineering — Kindle](https://www.amazon.com.br/dp/B0GZ4CYY6W)
- [Practical Claude Code — Engenharia de Contexto que Transforma seu Desenvolvimento — Landing Page](https://kenimoto.dev/pt/books/claude-code-mastery/)
- [Practical Claude Code — Kindle](https://www.amazon.com.br/dp/B0GZHB7J3T)

### Spanish (Español)

- [Convirtiendo LLMs de Mentirosos en Expertos — Ingeniería de Contexto en la Práctica — Landing Page](https://kenimoto.dev/es/books/context-engineering/)
- [Convirtiendo LLMs de Mentirosos en Expertos — Kindle](https://www.amazon.com.mx/dp/B0H12JR228)
- [Harness Engineering — De Usar IA a Controlar IA — Landing Page](https://kenimoto.dev/es/books/harness-engineering-guide/)
- [Harness Engineering — Kindle](https://www.amazon.com.mx/dp/B0GXSVHZ5G)
- [Practical Claude Code — La Ingeniería de Contexto que Transforma tu Desarrollo — Landing Page](https://kenimoto.dev/es/books/claude-code-mastery/)
- [Practical Claude Code — Kindle](https://www.amazon.com.mx/dp/B0GX3S7VGK)

## Blog Articles (462 articles, auto-generated)

### English

- [2.8x Traffic Spike: How to Identify Bot Traffic](https://kenimoto.dev/blog/crawler-three-types-user-agent-check/)
- [Bing Webmaster Tools vs Google: 787 queries](https://kenimoto.dev/blog/google-search-demand-suggest-echo-787-queries/)
- [Zenn noindex on Every Post: 8 Weeks Unnoticed](https://kenimoto.dev/blog/zenn-noindex-8-weeks-unnoticed/)
- [Harness Engineering Benchmark: 13.7pt Same Model](https://kenimoto.dev/blog/harness-vs-prompt-52-to-66/)
- [pgvector vs Qdrant vs Weaviate at 10k: No Cliff](https://kenimoto.dev/blog/pgvector-vs-qdrant-vs-weaviate-10k-agent-memories/)
- [I Measured AI Citation Half-Life for 90 Days: 6 of 10 Sources Vanish by Week 4](https://kenimoto.dev/blog/ai-citation-90-days-6-of-10-sources-vanish/)
- [Static Call Graphs Miss 61% of Invocations](https://kenimoto.dev/blog/static-call-graphs-miss-61-percent-runtime-traces-ai-code-review/)
- [AI Mode Cited My Portuguese 3.4× English. Japanese Also Beat English.](https://kenimoto.dev/blog/ai-mode-portuguese-vs-english-cross-language-llmo/)
- [Claude Skills vs Subagents in 2026: Which One To Reach For (7 Decision Rules)](https://kenimoto.dev/blog/claude-skills-vs-subagents-7-decision-rules/)
- [Claude Code vs ChatGPT Codex: 30-Day Cost by 7 Task Types](https://kenimoto.dev/blog/claude-code-vs-chatgpt-codex-30-day-cost-per-task-type/)
- [Spec-Driven Development with Claude Code: 3 Ways the Spec Itself Broke Us](https://kenimoto.dev/blog/spec-driven-development-claude-code-spec-broke-us/)
- [Codex CLI vs Claude Code: 7 Real Tasks, Same Repo, 31 Days Later](https://kenimoto.dev/blog/codex-cli-vs-claude-code-7-real-tasks-31-days/)
- [GitHub Copilot Agent Mode vs Claude Code: 8 Tasks, 31 Days, Real Bills](https://kenimoto.dev/blog/github-copilot-agent-mode-vs-claude-code-8-tasks-31-days/)
- [Claude Code vs Aider on Same 10 Tasks — One Finished 8, the Other 5](https://kenimoto.dev/blog/claude-code-vs-aider-10-tasks-eight-vs-five/)
- [Claude Code vs Qwen3-35B on RTX 4070: 34.6 tok/s Break-even](https://kenimoto.dev/blog/claude-code-vs-qwen3-35b-rtx-4070-benchmark/)
- [Karpathy's nanochat on a MacBook: End-to-End runcpu.sh Log from Tokenizer to Chat (2026)](https://kenimoto.dev/blog/karpathy-nanochat-macbook-runcpu-end-to-end-log/)
- [Claude Code Login: 2 Auth Layers vs setup-token](https://kenimoto.dev/blog/claude-code-two-layer-auth-setup-token/)
- [The Shai-Hulud npm attack: why signature verification, npm audit, and --ignore-scripts all failed](https://kenimoto.dev/blog/shai-hulud-npm-supply-chain-why-verification-failed/)
- [Claude Code Diagrams: 3 Conditions That Work](https://kenimoto.dev/blog/claude-code-shared-canvas-diagramming/)
- [Claude Code: The Operator's Field Guide](https://kenimoto.dev/blog/claude-code/)
- [llms.txt vs Claude Skills manifest: I tested 5 AI engines to see which file they actually read](https://kenimoto.dev/blog/llms-txt-vs-claude-skills-manifest-5-ai-engines/)
- [When Claude Code's Auto Mode Blocks Only Bash: Investigating the Safety Classifier Outage, Plus a Fail-Open-on-Outage Hook Design](https://kenimoto.dev/blog/claude-code-auto-mode-classifier-fail-open-hook/)
- [Chat Models Are Born in the Loss Mask — Reading nanochat's SFT](https://kenimoto.dev/blog/nanochat-sft-loss-mask-where-chat-models-are-born/)
- [Five Days of Tracking 346 Aftershocks From 90 km Away — and a Rematch With the 2016 Kumamoto Earthquake](https://kenimoto.dev/blog/uto-m71-aftershock-five-days-fukuoka/)
- [nanochat: $48 to Retrain GPT-2 (8×H100, 2026)](https://kenimoto.dev/blog/karpathy-nanochat-gpt-2-training-cost-43000-to-48-h100/)
- [JMA Earthquake Prediction Impossible: Day 8 Miss](https://kenimoto.dev/blog/earthquake-prediction-vs-aftershock-forecasting/)
- [llmoframework Audit on 30 Dev Blogs: 4 of Top 5 Fail Same 3 Checks](https://kenimoto.dev/blog/llmoframework-audit-30-dev-blogs-top-5-fail-same-3-checks/)
- [Cursor Composer vs Claude Code: 400k-Token Repo Benchmarked](https://kenimoto.dev/blog/cursor-composer-vs-claude-code-400k-token-repo/)
- [nanochat: I Retrained GPT-2 for $48 in 2026](https://kenimoto.dev/blog/karpathy-nanochat-43k-gpt2-8000-lines/)
- [PinchTab Only Shoots One Viewport: Teaching a 9.4k-Star Browser Bridge to Capture Full Pages](https://kenimoto.dev/blog/pinchtab-scroll-shot-full-page-screenshots/)
- [pixelshot Read One Tile of an 18,609px Wikipedia Page: the Lazy-Load Trap Under Visual RAG](https://kenimoto.dev/blog/pixelshot-1-tile-wikipedia-lazy-load-trap/)
- [The Machine Accent Travels: AI Text Is Rhythmically Monotone in 70/70 Cells Across 3 Languages](https://kenimoto.dev/blog/machine-accent-3-languages-70-70-cells/)
- [Article Schema Alone Didn't Make AI Recognize Me as the Author. The Entity Wiring That Did (in 4 JSON-LD Fields).](https://kenimoto.dev/blog/article-schema-alone-author-entity-4-json-ld-fields/)
- [OpenCut MCP: Drive a Video Editor from Claude Code (4 Traps)](https://kenimoto.dev/blog/opencut-classic-mcp-4-traps-editor-core-fork/)
- [MCP Server Audit: 4 Layers, A-F Grade](https://kenimoto.dev/blog/pre-flight-your-mcp-4-layers-scorecard/)
- [The Skill Eval Repo I Didn't Build: 107 SKILL.md Files, 6 Checks, 21 False Positives](https://kenimoto.dev/blog/skill-eval-repo-not-built-107-lint/)
- [Ship a product, get a support button for free: an edge-injected overlay on Cloudflare Workers](https://kenimoto.dev/blog/edge-injected-support-overlay-cloudflare-workers/)
- [FUNDING.yml alone won't show a Sponsor button: notes from auditing 42 repos](https://kenimoto.dev/blog/funding-yml-sponsor-button-42-repo-audit/)
- [historymap: one YAML file becomes a corporate-style product-history timeline](https://kenimoto.dev/blog/historymap-yaml-corporate-timeline-oss/)
- [Claude Code Made Me 40% Slower: 3 Places](https://kenimoto.dev/blog/claude-code-40-slower-3-places/)
- [The 7-Step Test That Told Me When to Switch From RAG to GraphRAG](https://kenimoto.dev/blog/graphrag-vs-rag-7-step-switch-test/)
- [I Ran Claude Code, Cursor, and Codex Side by Side for 31 Days. Here Is the Real Monthly Bill.](https://kenimoto.dev/blog/claude-code-cursor-codex-31-days-real-monthly-bill/)
- [Claude Code vs Cursor: 6 Tasks, 1 Uninstalled](https://kenimoto.dev/blog/claude-code-vs-cursor-6-tasks-uninstalled/)
- [Claude Code, Cursor, Codex: 340k Token Breakeven](https://kenimoto.dev/blog/parallel-agents-340k-tokens-breakeven-claude-cursor-codex/)
- [RTX 4070 llama.cpp --cpu-moe: 12.2 to 34.6 tok/s](https://kenimoto.dev/blog/rtx-4070-cpu-moe-flag-2-8x-tokens/)
- [Your AI Code Review Burns 80% of the Context Window on Files It Never Needed](https://kenimoto.dev/blog/ai-code-review-80-percent-context-waste/)
- [I Made Claude Code Review Only the Blast Radius — Token Bill Dropped 8-49x](https://kenimoto.dev/blog/claude-code-blast-radius-review-tokens-8-49x/)
- [Anthropic frontend-design skill: #F4F1EA Named](https://kenimoto.dev/blog/anthropic-frontend-design-skill-rewrite/)
- [Claude Code, Cursor, Codex: 412 Decisions/Day](https://kenimoto.dev/blog/claude-cursor-codex-parallel-decision-fatigue-400/)
- [Measuring AI Citation Half-Life: A 90-Day Methodology With 3 Real Decay Curves](https://kenimoto.dev/blog/measuring-ai-citation-half-life-90-day-methodology/)
- [Multi-Agent Decision Fatigue: I Counted 412 Micro-Choices a Day. The Harness Cut It to 38.](https://kenimoto.dev/blog/multi-agent-decision-fatigue-412-to-38/)
- [I Added 3 Numbers to One Paragraph. Perplexity Started Citing It in 11 Days.](https://kenimoto.dev/blog/perplexity-cited-3-numbers-11d/)
- [AI Mode Just Hit 1 Billion Users, and Opened a Local-Business LLMO Market Most Engineers Are Ignoring](https://kenimoto.dev/blog/ai-mode-billion-users-local-business-llmo/)
- [I Wired My Pages Into Topic Hubs, Not a Flat List: AI Citations Consolidated Onto 4 of Them](https://kenimoto.dev/blog/topic-hubs-ai-citations/)
- [AI Search Is Under 1% of My Traffic and 12% of My Signups. That's the LLMO Case I Actually Use.](https://kenimoto.dev/blog/llmo-roi-23x-conversion/)
- [GEO's +115% From Statistics Is Domain-Dependent: It Worked for My Tech Posts and Did Nothing for My How-To Pages](https://kenimoto.dev/blog/geo-stats-domain-dependent/)
- [AI Reads Your Chunks, Not Your Page: I Promoted 9 Sections from H3 to H2 and Watched Which Ones Got Quoted](https://kenimoto.dev/blog/ai-reads-chunks-not-pages/)
- [I Wired Claude Code to Real Hardware Over USB Serial. The MCP Tool Was the Easy Part.](https://kenimoto.dev/blog/claude-code-real-hardware-usb/)
- [I Checked What GPTBot Actually Sees on My JS-Rendered Pages. It Was an Empty `<div>`.](https://kenimoto.dev/blog/ai-crawlers-dont-execute-javascript-invisible-pages/)
- [AI Wrote 100 Passing Tests. Mutation Testing Says They Caught 58% of Real Bugs.](https://kenimoto.dev/blog/ai-100-tests-mutation-score/)
- [AI Finds Your Page Three Ways. I Published the Same Fact in All Three and Timed Which Reached AI First.](https://kenimoto.dev/blog/ai-finds-your-page-three-ways/)
- [Claude Code Pricing vs API vs Local GPU in 2026](https://kenimoto.dev/blog/ai-agent-monthly-cost-api-subscription-local-breakeven/)
- [Multilingual LLMO: 4 Languages, Wrong Citation](https://kenimoto.dev/blog/ai-cites-wrong-language-version-multilingual-llmo/)
- [I Rewrote 12 Pages to Answer the Question in the First Sentence. AI Started Quoting 7 of Them.](https://kenimoto.dev/blog/answer-first-7-of-12-cited/)
- [I Rank #1 on Google. On Brave I'm Page 5. My Own AI Agents Can't Find Me.](https://kenimoto.dev/blog/brave-invisible-to-ai-agents/)
- [Perplexity Citations Exploded After I Changed 3 Things. Only 1 Was Schema.](https://kenimoto.dev/blog/perplexity-3-changes-1-schema/)
- [I Gave Every Page on My Site a .md Twin. The AI Fetchers Stopped Guessing](https://kenimoto.dev/blog/every-page-md-twin-llmo/)
- [My Best Page Went Stale in a Month: Why AI Search Rewards Freshness, Not Just Schema](https://kenimoto.dev/blog/ai-search-rewards-freshness/)
- [AI Search Splits Your One Question Into Six. My Pages Answered None of Them.](https://kenimoto.dev/blog/query-fanout-ai-citations/)
- [I Stopped Adding Context to My Agent and Pruned Tool Outputs Instead — My 3-Hour Task Stopped Forgetting Its Own Plan](https://kenimoto.dev/blog/stopped-adding-context-pruned-tool-outputs-accuracy-returned/)
- [LLM Citation Decay: 9 Weeks, Peak to 46 Percent](https://kenimoto.dev/blog/ai-citations-half-life-decay/)
- [Blast Radius: The File That Breaks Is 2 Hops Out](https://kenimoto.dev/blog/codebase-graph-two-hops/)
- [Link-less Brand Mentions Beat Backlinks for AI Visibility — I Read the Ahrefs 75,000-Brand Study So You Don't Have To](https://kenimoto.dev/blog/mentions-beat-backlinks-ai/)
- [Your Page Rank Is Invisible to AI — Only Your Passages Get Cited](https://kenimoto.dev/blog/passage-rank-beats-page-rank-ai-citations/)
- [I Crosspost to 4 Platforms with rel=canonical Pointing Home. AI Search Still Picks the Copy.](https://kenimoto.dev/blog/crosspost-canonical-ai-picks-the-copy/)
- [Claude Code Skills Cost Tokens Even When They Don't Fire. I Measured 5 Skills Across 7 Hours. The Bill Was 18%.](https://kenimoto.dev/blog/skills-loaded-3-never-fired-18/)
- [I Cron-Scheduled 7 AI Agents. 2 Silently Failed for 18 Days. Tracing Wouldn't Have Caught It.](https://kenimoto.dev/blog/seven-cron-agents-18d-silent/)
- [I Ran 3 Claude Code Sessions in Parallel for 8 Hours. They Overwrote Each Other's Context Twice.](https://kenimoto.dev/blog/three-claude-sessions-parallel-8h-context-overwrite/)
- [I Asked 5 AI Search Engines to Cite My Own Blog. Only 3 of 31 Articles Showed Up.](https://kenimoto.dev/blog/five-ai-engines-cite-my-blog-three-of-thirty-one/)
- [I Added 11 JSON-LD Schemas. Three Months Later, Only 3 Showed Up in AI Citations.](https://kenimoto.dev/blog/11-json-ld-3-cited-by-ai/)
- [I Refactored 100 Functions With Claude. 7 Got Slower in Production.](https://kenimoto.dev/blog/claude-refactor-100-functions-7-slower-production/)
- [Claude Code Hooks for TDD: 4 of 10 to 9 of 10](https://kenimoto.dev/blog/claude-code-tdd-test-after-code-six-of-ten/)
- [I Added a 4th Agent That Audits My Other Agents. It Caught My Strategist Procrastinating for 3 Weeks.](https://kenimoto.dev/blog/evolver-fourth-agent-caught-strategist-procrastinating/)
- [I Translated My Blog Into 4 Languages. Portuguese Got Nearly 4× the Traffic of English.](https://kenimoto.dev/blog/four-languages-thirty-days-portuguese-four-x-traffic/)
- [TRM 8,337% LLMO Playbook: 1 of 4 Pillars Worked](https://kenimoto.dev/blog/trm-8337-percent-llmo-pillars-indie-test/)
- [Claude Said 'You're Absolutely Right!' 47 Times Last Week. I Was Only Right 11 Times. Claude Was Wrong 36.](https://kenimoto.dev/blog/claude-sycophancy-47-times-measured/)
- [I Plugged the Same Site Into 7 AI-Citation Trackers. They Reported 7 Different Numbers.](https://kenimoto.dev/blog/seven-ai-citation-trackers-seven-different-numbers/)
- [The 5 AI Crawlers That Hit My Sites Most in 30 Days — What Their Logs Told Me About LLMO](https://kenimoto.dev/blog/five-ai-crawlers-hit-my-site-30-days/)
- [I Plugged Claude into a Chaos Engineering MCP Server. It Killed Staging 4 Times Before Finding a Bug We'd Missed for 6 Months.](https://kenimoto.dev/blog/claude-chaos-engineering-mcp-killed-staging-4-times/)
- [I Caught Claude Hiding My Bug 3 Times in a Row. Then I Turned 10 Debugging Habits Into Prompts.](https://kenimoto.dev/blog/claude-hid-my-bug-three-times-ten-debugging-prompts/)
- [I Gave My Strategist Agent WebSearch. 5 Topics Took 20 Minutes. Splitting It Into 3 Made It 3.](https://kenimoto.dev/blog/three-role-separation-observer-strategist-marketer/)
- [I Benchmarked 5 Voice AI Stacks. Only 2 Stayed Under 300ms.](https://kenimoto.dev/blog/five-voice-ai-stacks-only-two-under-300ms/)
- [Claude Code Sub-Agents Disagreed on 40% of My PR](https://kenimoto.dev/blog/three-sub-agents-reviewed-same-pr-40-percent-disagreement/)
- [llms.txt Audited: 30 Files, 5 Anti-Patterns](https://kenimoto.dev/blog/30-llms-txt-files-5-anti-patterns-already-forming/)
- [OpenClaw Hit 250K Stars Faster Than React. I Spent a Day Switching From Claude Code](https://kenimoto.dev/blog/openclaw-vs-claude-code-24h/)
- [I Refused to Write Specs Until Claude Code Generated Wrong Code Three Times](https://kenimoto.dev/blog/spec-driven-development-claude-code-three-failures/)
- [I Let My Claude Code Agent Run for 24 Hours. The $400 Bill Was the Least Scary Part.](https://kenimoto.dev/blog/autonomous-agent-24-hours-security-lessons/)
- [The og:type Bug Three of My Astro Sites Quietly Shipped](https://kenimoto.dev/blog/og-type-double-emit-three-astro-sites/)
- [I Stacked 4 More Context Layers on Top of RAG. The Improvement Was 12%.](https://kenimoto.dev/blog/full-context-engineering-rag-80-percent/)
- [arXiv 2603.25723 \"Natural-Language Agent Harnesses\": 4 Patterns + 3 Anti-Patterns After 12 Weeks in Prod](https://kenimoto.dev/blog/natural-language-agent-harnesses-arxiv/)
- [Claude Code Skills vs Commands: 47 Prompts](https://kenimoto.dev/blog/claude-code-skills-reusable-workflow-pattern/)
- [Your New Domain's First Week of GA4 Is a Lie: 4 Days of Raw Data from kaoriq.com's Launch](https://kenimoto.dev/blog/new-domain-first-week-ga4-is-a-lie/)
- [ChatGPT Codex vs Claude Code: 47 PRs, $297 Bill, Aug 2026](https://kenimoto.dev/blog/claude-code-vs-chatgpt-codex-official-agents/)
- [One Question, Five AI Search Engines, Five Different Answers](https://kenimoto.dev/blog/five-ai-search-engines-architecture-llmo/)
- [Is AI Actually Citing Your Site? How to Measure What Google Rankings Can't](https://kenimoto.dev/blog/measure-ai-citations-llmo-kpi/)
- [Generative Engine Optimization: 3 of 9 Worked](https://kenimoto.dev/blog/geo-princeton-study-9-ways-ai-cites-you/)
- [9 Bugs in My AI Pipeline: None Were the AI's Fault](https://kenimoto.dev/blog/9-bugs-in-my-ai-pipeline/)
- [Claude Haiku + RAG Beat Sonnet 11.8 to 5.3](https://kenimoto.dev/blog/cheap-model-won-context-beats-parameters/)
- [llms.txt: The File That Decides Whether AI Can Find Your Site](https://kenimoto.dev/blog/llms-txt-ai-find-your-site/)
- [Building an Autonomous Content Pipeline with Claude Code](https://kenimoto.dev/blog/hello-world/)

### Japanese (日本語)

- [音声AI遅延の3つの崖 200/300/800ms](https://kenimoto.dev/ja/blog/voice-ai-three-cliffs-200-300-800ms/)
- [Claude Code Skills 発火率10%](https://kenimoto.dev/ja/blog/claude-code-skills-firing-10-of-101/)
- [AIコードレビュー、Diffで依存を見逃す実測](https://kenimoto.dev/ja/blog/ai-code-review-diff-only-dependency-blindspot/)
- [MCP tool descriptionの穴7本](https://kenimoto.dev/ja/blog/mcp-description-lint-slipped-invariant-poc/)
- [sudoers引数制限の4段階:*は空白を跨ぐ](https://kenimoto.dev/ja/blog/sudoers-narrow-args-4-steps-wildcard-trap/)
- [アクセス2.8倍の正体、クローラー3種の見分け方](https://kenimoto.dev/ja/blog/crawler-three-types-user-agent-check/)
- [検索ボリューム調べ方: サジェスト787件実測](https://kenimoto.dev/ja/blog/google-search-demand-suggest-echo-787-queries/)
- [音声AIの525ms壁を300ms未満にする3手法](https://kenimoto.dev/ja/blog/voice-ai-cascade-525ms-streaming-3-methods/)
- [TailscaleでPiにCIデプロイ:権限3層](https://kenimoto.dev/ja/blog/tailscale-github-actions-pi-deploy-3-layer-permissions/)
- [Claude Codeサブエージェントの3階層権限](https://kenimoto.dev/ja/blog/claude-code-sub-agent-3-tier-permissions/)
- [Zenn全記事noindex、気づくまで8週間](https://kenimoto.dev/ja/blog/zenn-noindex-8-weeks-unnoticed/)
- [PRラベルnit/must機械分別で48h→24h](https://kenimoto.dev/ja/blog/conventional-comments-nit-must-48h-24h-ja/)
- [KDP 7枠をBing需要で選んで外した実測](https://kenimoto.dev/ja/blog/kdp-keyword-bing-demand-miss-amazon-suggest/)
- [Google Play Booksの20%公開義務](https://kenimoto.dev/ja/blog/play-books-20-percent-preview-and-ai-training/)
- [voice AI TTFBを100msに詰めた実測](https://kenimoto.dev/ja/blog/voice-ai-ttfb-100ms-300ms-wall/)
- [GraphRAGでLinkedIn 28.6%短縮](https://kenimoto.dev/ja/blog/graphrag-linkedin-28-6/)
- [Whisper文字起こしを使わない歌詞同期: 実測誤差0.12秒](https://kenimoto.dev/ja/blog/lyrics-timing-forced-alignment-not-asr/)
- [PINとパスワードの違いは桁数ではない: 6桁で足りる理由](https://kenimoto.dev/ja/blog/pin-vs-password-not-about-length/)
- [技術書116冊のA+コンテンツを数えたら、個人出版は53冊中11冊しか入れていなかった](https://kenimoto.dev/ja/blog/aplus-content-116-technical-books/)
- [「浮き彫りにする」はAIの180記事に1回も出てこなかった — 日本語AI頻出語の実測1位は「適切な」](https://kenimoto.dev/ja/blog/ukibori-zero-in-180-ai-vocabulary-measured/)
- [MCP承認ゲートを3層に分けた: allow/ask/denyの運用ログ](https://kenimoto.dev/ja/blog/mcp-approval-gate-3-layers-allow-ask-deny/)
- [Claude Code Skills に入れると壊れる処理3種類と、分離先の選び方](https://kenimoto.dev/ja/blog/claude-code-skills-mcp-hook-boundary/)
- [音声AIの barge-in 実装で認識精度 22% 減 — VAD 閾値 4 段階実測](https://kenimoto.dev/ja/blog/voice-ai-barge-in-22-percent-accuracy-drop-vad-4-thresholds/)
- [「AI経由の流入が増えた」は自分で作れる。導線を入れる前に測り方を分けた](https://kenimoto.dev/ja/blog/ai-referral-self-inflation-measure-first/)
- [AI時代の新バイアス3選 — 自動化バイアスで見逃したバグと3ヶ月の実測対策](https://kenimoto.dev/ja/blog/ai-automation-bias-3-months-3-bugs/)
- [採譜は動いた。伴奏にはならなかった — OSSでカラオケ音源を作れなかった記録](https://kenimoto.dev/ja/blog/karaoke-oss-transcription-worked-arrangement-missing/)
- [CLAUDE.mdの作者は2行、実践者は100行 — 情報密度の非対称論](https://kenimoto.dev/ja/blog/claude-md-2-vs-100-lines-jouhou-mitsudo/)
- [MCP Sampling権限昇格を止める防御実装3パターン - Claude Desktopで実測](https://kenimoto.dev/ja/blog/mcp-sampling-defense-3-patterns-claude-desktop-jissoku/)
- [OpenVPNはなぜ.ovpnファイルとパスワードだけでVPNに繋がるのか](https://kenimoto.dev/ja/blog/openvpn-config-file-and-authentication-flow/)
- [Tailscale WireGuard比較 3層](https://kenimoto.dev/ja/blog/wireguard-openvpn-tailscale-protocol-layers/)
- [Claude Codeの認証は2層構造になっている。setup-tokenとloginは別物だ](https://kenimoto.dev/ja/blog/claude-code-two-layer-auth-setup-token/)
- [npmサプライチェーン攻撃Shai-Hulud：署名検証もSBOMも無意味だった理由](https://kenimoto.dev/ja/blog/shai-hulud-npm-supply-chain-why-verification-failed/)
- [Claude Codeに図を「描かせる」のをやめて「一緒に描く」に変えたら、図解の意味が変わった](https://kenimoto.dev/ja/blog/claude-code-shared-canvas-diagramming/)
- [Claude Code 実戦運用ガイド ― エージェントを2年回して分かった8つの勘所](https://kenimoto.dev/ja/blog/claude-code/)
- [静的解析コードKGが落とす61%の動的呼び出し](https://kenimoto.dev/ja/blog/kg-61-percent-miss-dynamic-call-payment-rollback/)
- [RTX 4070 一枚で BGM・効果音・画像を全部ローカル生成する — ACE-Step 1.5 と MOSS-SoundEffect v2.0 を一晩で実戦投入した話](https://kenimoto.dev/ja/blog/rtx4070-local-bgm-sfx-ace-step-moss-soundeffect/)
- [Skill・Hook・MCPで同じ機能を3実装した結果、月額トークン費が家賃並みに割れた](https://kenimoto.dev/ja/blog/skill-hook-mcp-3-jissou-getsugaku-yachin/)
- [AIコードレビューは説得の心理学:3つの言い回しでmerge速度が1.5倍になった話](https://kenimoto.dev/ja/blog/ai-code-review-persuasion-psychology-1-5x-merge-speed/)
- [Claude Code の auto mode で Bash だけが止まる — safety classifier 障害の調査と「故障時だけ fail-open」hook 案](https://kenimoto.dev/ja/blog/claude-code-auto-mode-classifier-fail-open-hook/)
- [LLMOだけでは読まれない。llms.txt完備のサイトで、AI経由の流入は3%だった](https://kenimoto.dev/ja/blog/llmo-only-3-percent-ai-referral-jissoku/)
- [チャットモデルは損失マスクで生まれる — nanochatのSFTを「採点表」として読む](https://kenimoto.dev/ja/blog/nanochat-sft-loss-mask-chat-model/)
- [Tauri v2 を「ガワ」として使う — WSL2 mirrored networking で Node.js サーバをそのまま Windows デスクトップに映す](https://kenimoto.dev/ja/blog/tauri-v2-gawa-wsl-mirrored-networking/)
- [熊本M7.1の余震346件を5日間追った — 福岡に届いたのは34件、そして10年前との答え合わせ](https://kenimoto.dev/ja/blog/uto-m71-yoshin-346-fukuoka-34-jissoku/)
- [地震予知はなぜ無理で、余震予測はなぜできるのか — 能登の余震130件で実測](https://kenimoto.dev/ja/blog/jishin-yochi-vs-yoshin-yosoku-jissoku/)
- [Shai-Hulud事件でCLAUDE.mdの書き方が変わった話](https://kenimoto.dev/ja/blog/shai-hulud-claude-md-3-removed-5-added/)
- [Claude Code MCPでハードウェア制御——UART/I2C/SPI/RS-485/CAN 5プロトコル実測](https://kenimoto.dev/ja/blog/claude-code-mcp-hardware-5-jissoku/)
- [MEO代行費5万円をClaudeで自前化、3業種プロンプト実物と月3万円分解](https://kenimoto.dev/ja/blog/meo-daiko-claude-jimae-3gyoushu/)
- [MCP OWASP Top 10全10項目、Claude Desktop標準設定で何が守れて何が抜けるか](https://kenimoto.dev/ja/blog/mcp-owasp-top10-claude-desktop-defense-map-2026-07/)
- [MCP Sampling 権限昇格 — Claude Desktop で通る3経路 (7攻撃検証)](https://kenimoto.dev/ja/blog/mcp-sampling-privilege-escalation-claude-desktop-3-paths/)
- [「リズムの単調化は日本語だけか」を英語・ポルトガル語×7モデルで検証したら、70セル全部でAIは単調だった](https://kenimoto.dev/ja/blog/machine-accent-3-languages-70-70-cells/)
- [ローカルLLMを2フラグで34.6 tok/s](https://kenimoto.dev/ja/blog/rtx4070-qwen35b-2flags-kv-quant-benchmark/)
- [「AI臭は語彙よりリズムに出る」を7モデルで検証したら、判別はAUC 0.998で語彙の圧勝だった](https://kenimoto.dev/ja/blog/japanese-ai-smell-vocabulary-vs-rhythm-fingerprints/)
- [nanochat: GPT-2訓練の2026年値段](https://kenimoto.dev/ja/blog/nanochat-gpt2-training-cost-2026/)
- [論文の実測分布で校正したリズム計測CLI「rhythm-lens」を公開して、自分の記事に当てたら⚠2つ食らった](https://kenimoto.dev/ja/blog/rhythm-lens-cli-own-article-2-warnings/)
- [1Passwordを値上げで諦めた私が、SOPS+ageで.envの複数PC共有に着地した話](https://kenimoto.dev/ja/blog/sops-age-multi-pc-env-envelope-model/)
- [複数アカウントで気づいた Claude Code の2ファイル認証 (credentials.json + ~/.claude.json)](https://kenimoto.dev/ja/blog/claude-code-two-file-auth-multi-account/)
- [MCPのTool descriptionはプロンプトインジェクションの窓 - 公開5個で試した3種](https://kenimoto.dev/ja/blog/mcp-tool-description-injection-5-servers-3-patterns/)
- [Strangler Fig で React/FastAPI 段階移行](https://kenimoto.dev/ja/blog/strangler-fig-react-fastapi-4phase/)
- [YAML→PNG を LLM に任せる — historymap に MCP サーバーを足した記録](https://kenimoto.dev/ja/blog/historymap-mcp-server-yaml-to-png/)
- [MCPのファイル添付7サービス実測: 対応ゼロ](https://kenimoto.dev/ja/blog/mcp-file-upload-7-services-benchmarked-full-support-zero/)
- [X Premium ¥980 では Grok を CLI から叩けない — 403 の理由と $25 credit で始める最短路](https://kenimoto.dev/ja/blog/x-premium-980-grok-cli-403-and-cheapest-path/)
- [AGENTS.md と CLAUDE.md の違いは「専用機か汎用機か」ではない — 3ヶ月使い分けて分かった実践設計の分岐点](https://kenimoto.dev/ja/blog/agents-md-claude-md-practical-design-branch/)
- [自作LLMOサイト、日本人1人運営なのに中国からのアクセスが45%になっていた — cn.bing で自然発生的にランクした話](https://kenimoto.dev/ja/blog/llmoframework-cn-bing-china-45pct-organic-discovery/)
- [OpenCut MCP: Claude Code から動画エディタを操作する（4つの罠）](https://kenimoto.dev/ja/blog/opencut-classic-mcp-4-traps-editor-core-fork/)
- [OpenCutセットアップ難民は、まずopencut.appを開いてください — ローカル起動の3つの罠](https://kenimoto.dev/ja/blog/opencut-setup-nanmin-webapp-first-3-traps/)
- [MCPサーバーを公開する前のpre-flight — 4層で採点する mcp-scorecard の設計](https://kenimoto.dev/ja/blog/pre-flight-your-mcp-4-layers-scorecard/)
- [AIエージェントのハーネスを構成する6要素 — 10分でCLAUDE.mdを棚卸しするチェックリスト](https://kenimoto.dev/ja/blog/harness-6-components-claude-md-10min-checklist/)
- [107本のSKILL.mdをlintしたら、稼働中のskillが21本「壊れてる」と言われた](https://kenimoto.dev/ja/blog/skill-eval-two-layer-107-lint/)
- [プロダクトを出すたび支援導線が勝手に付く: Cloudflare Workersのエッジ注入オーバーレイ](https://kenimoto.dev/ja/blog/edge-injected-support-overlay-cloudflare-workers/)
- [FUNDING.ymlを置いてもSponsorボタンは出ない: 42リポジトリを棚卸しした](https://kenimoto.dev/ja/blog/github-sponsor-button-42-repo-audit/)
- [YAMLを1枚書くと企業サイト風の年表ページになるOSS「historymap」を公開した](https://kenimoto.dev/ja/blog/historymap-yaml-timeline-oss/)
- [自己進化エージェントを3ヶ月動かしたら2回ロールバックした](https://kenimoto.dev/ja/blog/self-evolving-agent-3-months-2-rollbacks/)
- [MCPで私が踏んだ地雷7つ — 実装と実測ログ](https://kenimoto.dev/ja/blog/mcp-7-mines-implementation-log/)
- [MCPサーバーの権限昇格を7パターン実演、うち3つはClaude Desktopの標準設定で通った](https://kenimoto.dev/ja/blog/mcp-owasp-privilege-escalation-7-attacks-claude-desktop/)
- [リリース作業の途中で、作っていたプラグインに67%警告を鳴らされた: compact-ops を公開しました](https://kenimoto.dev/ja/blog/compact-ops-dogfooding-67-percent-warn/)
- [Claude Code 3セッション並走で8時間ぶん消えた: あの日から追加した5つのハーネス制約](https://kenimoto.dev/ja/blog/three-sessions-5-harness-rules/)
- [Knowledge Graphを7ステップで作る前に、AIエージェントが「見えていない」3層を可視化する](https://kenimoto.dev/ja/blog/knowledge-graph-3-blind-layers-before-7-steps/)
- [llms.txtとJSON-LDが噛み合っていない5パターン — 30ファイルを整合チェックして分かったこと](https://kenimoto.dev/ja/blog/llms-txt-json-ld-integrity-5-patterns/)
- [LLMO 総合ガイド ― AI 検索に引用される技術を、30ヶ月分の計測から組み立てる](https://kenimoto.dev/ja/blog/llmo/)
- [Ollama 12.2 tok/s が34.6 tok/s になった: RTX 4070 で Qwen3 35B を2.8倍速にする1コマンド](https://kenimoto.dev/ja/blog/ollama-qwen3-35b-rtx-4070-cpu-moe-2-8x/)
- [Claude Code・Cursor・Codexを1日並列で走らせて、判断回数を数えたら412回だった](https://kenimoto.dev/ja/blog/claude-cursor-codex-heiretsu-handan-412kai/)
- [Claude Code・Cursor・Codex併用で1日の判断を400回から40回にした、ハーネス3層設計の実測](https://kenimoto.dev/ja/blog/harness-3-layers-judgment-count-one-tenth/)
- [GraphRAGはRDFかプロパティグラフか3問](https://kenimoto.dev/ja/blog/graphrag-2026-rdf-vs-property-graph-3-questions/)
- [Sonnet 4は確信度4.2で嘘をついた、Haikuは「知らない」と答えた — 大きいモデルほど自信を持って幻覚する](https://kenimoto.dev/ja/blog/sonnet-4-confidence-4-2-haiku/)
- [Claude Code の Plan モードで「朝に一日を設計」したら、手戻りが半分になった](https://kenimoto.dev/ja/blog/plan-mode-morning-rework/)
- [却下ログは「学習信号」だった — 自己進化ハーネスが strategy.md に書けない暗黙知を拾う仕組み](https://kenimoto.dev/ja/blog/rejection-log-learning-signal/)
- [「著者エンティティを実装した」と書いた私が、自分の239ページを監査したら配線が途中で切れていた話](https://kenimoto.dev/ja/blog/author-entity-self-audit/)
- [賢いモデルが来るほど、ハーネスは「捨てる」設計になる — Anthropicがマルチからシングルに戻した理由](https://kenimoto.dev/ja/blog/smarter-model-thinner-harness/)
- [マルチエージェントは並列化で速くなる、は幻想だった — 調整コストはエージェント数の二乗で増える](https://kenimoto.dev/ja/blog/multi-agent-coordination-n2/)
- [Figma MakeとClaude Codeで、デザイナー不在のUIを作る — コードからデザインを逆生成する時代](https://kenimoto.dev/ja/blog/figma-make-claude-code-design/)
- [記事にArticle schemaを入れてもAIは私を著者と認識しなかった。著者エンティティを実装して引用の帰属が変わった話](https://kenimoto.dev/ja/blog/author-entity-ai-citation/)
- [GPTBotに私のJSレンダリングページを見せたら、中身は空の`<div>`だった](https://kenimoto.dev/ja/blog/ai-crawler-javascript-jikkou-shinai-csr-invisible/)
- [LLMO対応はテストしないと静かに壊れる：AIクローラーの可読性をPlaywrightでCI検証する](https://kenimoto.dev/ja/blog/llmo-playwright-ci-test/)
- [賢いモデルほど、上手に嘘をつく](https://kenimoto.dev/ja/blog/bigger-models-lie-better/)
- [「いい感じに作って」はなぜ5回目で崩壊するのか -- コードの前に書く6つの定義書と3つの記憶ファイル](https://kenimoto.dev/ja/blog/document-first-claude-code-6-docs-3-memory-files/)
- [AIエージェントが「空気を読めない」本当の理由:常識ナレッジグラフという欠けたレイヤー](https://kenimoto.dev/ja/blog/agent-commonsense-knowledge-graph-missing-layer/)
- [接続したMCPサーバーの38%は認証なしだった — 60日で30件のCVEが出たMCPセキュリティの現実](https://kenimoto.dev/ja/blog/mcp-38-percent-no-auth/)
- [43秒の異常が24時間の障害になった: メトリクスの急変と漸増を見分ける時系列デバッグ](https://kenimoto.dev/ja/blog/spike-vs-gradient-time-series-debug/)
- [AIエージェントの予測を1つの数字で信じるな — Simulatorは点推定でなく確率分布を返すべき](https://kenimoto.dev/ja/blog/simulator-distribution-not-point-estimate/)
- [サブエージェントにメインの記憶を渡すのは事故だった：7ファイルのうち4つは遮断すべき](https://kenimoto.dev/ja/blog/sub-agent-memory-isolation/)
- [バリデータがバグっていた — 「安全です」と報告するシステム自身が壊れているとき](https://kenimoto.dev/ja/blog/validator-was-the-bug-grey-failure/)
- [「とりあえず全部許可」でClaude Codeを動かすと、.envの秘密がそのままAnthropicに渡る話](https://kenimoto.dev/ja/blog/claude-code-deny-rules-env/)
- [店舗LLMOの答えはllms.txtでなくGBPだった — 私が綺麗に書いたファイルを、AIは一度も引用しなかった](https://kenimoto.dev/ja/blog/llmo-local-business-gbp-not-llmstxt/)
- [Claude Codeに渡すコンテキストを足すのをやめた — ツール出力を間引いたら長時間タスクの精度が戻った](https://kenimoto.dev/ja/blog/stopped-adding-context-pruning-recovered-accuracy/)
- [「稼働率99.5%」と「5,000件の決済失敗」は同じ事実 — 障害報告のフレーミングが緊急度を反転させる](https://kenimoto.dev/ja/blog/incident-framing-99-percent/)
- [午後3時、あなたのコードレビュー承認率はほぼ0%になる：決断疲労の科学とAIエージェント時代の処方箋](https://kenimoto.dev/ja/blog/decision-fatigue-3pm-code-review-approval/)
- [ページ単位のSEOはAIに見えていない: 引用されるのは「パッセージ」だった](https://kenimoto.dev/ja/blog/passage-design-ai-citation-not-page-rank/)
- [Claude Codeを/clearせず9時間動かした日、コンテキストはどこで腐り始めたのか](https://kenimoto.dev/ja/blog/claude-code-9h-context-rot-token/)
- [Claude Code Skills は発火しなくてもトークンを食う — 5 Skills × 7時間セッションで「使われなかった」3つが18%食べていた](https://kenimoto.dev/ja/blog/skills-loaded-3-never-fired-18/)
- [AIエージェントを7本cronで毎日回したら、2本が18日間沈黙していた — observabilityでは拾えず、exit-code契約で拾えた話](https://kenimoto.dev/ja/blog/seven-cron-agents-18d-silent/)
- [Claude Codeを3セッション並列で8時間動かしたら、2回コンテキストを上書きしあった](https://kenimoto.dev/ja/blog/three-claude-sessions-parallel-8h-context-overwrite/)
- [5つのAI検索に『kenimoto.devを引用して』と頼んだ。記事31本のうち出てきたのは3本だった](https://kenimoto.dev/ja/blog/five-ai-engines-cite-my-blog-three-of-thirty-one/)
- [JSON-LDを11スキーマ入れた。3ヶ月測ったら、AIが拾っていたのは3つだけだった](https://kenimoto.dev/ja/blog/json-ld-11-only-3-cited/)
- [コードレビューを6段階にしたら、AIと人間の分業が見えた](https://kenimoto.dev/ja/blog/code-review-6-stages-ai-human-boundary/)
- [robots.txtにAIクローラー13個の個別ルールを書いた。30日後、守ったのは3つだけだった](https://kenimoto.dev/ja/blog/robots-txt-ai-crawler-rules-30-days-only-3-followed/)
- [AI Overviewsに載るための4条件を30日試した。効いたのは1つだけだった](https://kenimoto.dev/ja/blog/ai-overviews-4-conditions-30-days-only-one-worked/)
- [他のエージェントを監査する4層目を足したら、Strategistが3週間サボっていたことが発覚した話](https://kenimoto.dev/ja/blog/evolver-4-layer-strategist-procrastination-audit/)
- [ブログを4言語化したら、ポルトガル語版だけ流入が約4倍になった - 22日分の生データ](https://kenimoto.dev/ja/blog/4-languages-30-days-portuguese-4x-traffic/)
- [Claude Code Skills を10個書いたら、4個に統合された — Reusable Pattern が回り始める瞬間と回らない瞬間](https://kenimoto.dev/ja/blog/claude-code-skills-10-to-4-integration-pattern/)
- [Claudeに先週47回『おっしゃる通りです』と言われた。そのうち11回は私が、36回はClaudeが間違っていた。](https://kenimoto.dev/ja/blog/claude-osshatoori-47kai-sycophancy-jissoku/)
- [同じサイトを7つのAI引用トラッカーに入れたら、7つとも違う数字を返してきた](https://kenimoto.dev/ja/blog/seven-ai-citation-trackers-different-numbers/)
- [30日間サーバーログを見続けて分かった、私のサイトを最も叩いた5つのAIクローラー - そこから読めるLLMOシグナル](https://kenimoto.dev/ja/blog/five-ai-crawlers-30days-server-log/)
- [ClaudeをカオスエンジニアリングのMCPサーバーに繋いだら、ステージングを4回殺した — 6ヶ月見逃していた本番バグを見つけた話](https://kenimoto.dev/ja/blog/claude-chaos-engineering-mcp-staging-4-kai-koroshita/)
- [Claudeが3回連続でバグを「隠す修正」を出してきた話 — デバッグ10の技法をプロンプトに翻訳する](https://kenimoto.dev/ja/blog/claude-bug-kakushi-debug-10-techniques-prompt/)
- [Strategist に WebSearch を持たせたら 5テーマ選びに 20分かかった - Observer / Strategist / Marketer 3役分離で 3分にした話](https://kenimoto.dev/ja/blog/observer-strategist-marketer-3-yaku-bunri/)
- [音声AIスタックを5つ実測した。300msの壁を越えられたのは2つだけだった](https://kenimoto.dev/ja/blog/voice-ai-5-stacks-only-two-under-300ms/)
- [3人のサブエージェントに同じPRを見せたら、4割の指摘で意見が割れた話](https://kenimoto.dev/ja/blog/three-sub-agents-pr-review-40-percent-disagreement/)
- [他社の llms.txt を30個監査したら、すでに5つのアンチパターンが形になっていた](https://kenimoto.dev/ja/blog/llms-txt-audit-30-files-5-anti-patterns/)
- [SEOが壊れる日 — 私のAIエージェントは、もうGoogleを見ていなかった](https://kenimoto.dev/ja/blog/llmo-three-paths-introduction/)
- [CLAUDE.md は結局 Context Engineering を1ファイルに凝縮したものだった](https://kenimoto.dev/ja/blog/claude-md-context-engineering-practice/)
- [同じ質問なのに、LLMから5つの違う回答が返ってきた — Context Engineering 入門](https://kenimoto.dev/ja/blog/context-engineering-introduction-five-strategies/)
- [構造化データ11スキーマをLLM向けに統合](https://kenimoto.dev/ja/blog/json-ld-11-schemas-llm-understanding/)
- [ChatGPTからのアクセスは、GA4にどう映るのか - LLMOを数値で把握する3つの方法](https://kenimoto.dev/ja/blog/llmo-measurement-3-methods/)
- [LLMO最小実装は2ファイル15分で終わる](https://kenimoto.dev/ja/blog/llmo-minimum-implementation-llms-txt-json-ld/)
- [新規ドメインの最初の1週間、GA4は嘘をつく: kaoriq.com立ち上げ4日間の生データ](https://kenimoto.dev/ja/blog/new-domain-first-week-ga4-is-a-lie/)
- [ChatGPTからの流入を90日で8,337%増やした実装: TRMが採用した4つのLLMO戦略柱](https://kenimoto.dev/ja/blog/llmo-case-studies-trm-8337-percent/)
- [AIエージェントは月いくらかかるのか -- API・サブスク・ローカルの損益分岐点](https://kenimoto.dev/ja/blog/ai-agent-cost-structure-breakeven/)
- [Claude Code Hooks v2 — 「お願い」を「プログラム」に変える25のイベント](https://kenimoto.dev/ja/blog/claude-code-hooks-v2-25-events/)
- [Claude CodeのSub-agent設計 — 1セッションで専門家チームを使い分ける](https://kenimoto.dev/ja/blog/claude-code-sub-agent-design/)
- [AIパイプラインが壊れる9つの理由 -- 全部AIの外側だった](https://kenimoto.dev/ja/blog/9-bugs-in-my-ai-pipeline/)
- [LLMに「引用したい」と思わせるコンテンツ設計 — Microsoftが明かした3原則](https://kenimoto.dev/ja/blog/llm-content-design-microsoft-3-principles/)
- [MCPサーバーに接続しただけでトークンが消える — 4サービスで実測した](https://kenimoto.dev/ja/blog/mcp-token-cost-measurement/)
- [Claude Codeで自律型コンテンツパイプラインを構築した話](https://kenimoto.dev/ja/blog/hello-world/)

### Portuguese (Português)

- [3 armadilhas cognitivas em debugging: 2h → 15min](https://kenimoto.dev/pt/blog/3-armadilhas-cognitivas-debugging-2h-15min/)
- [5 vieses que decidem sua stack antes do teste](https://kenimoto.dev/pt/blog/5-vieses-que-decidem-sua-stack-antes-do-teste/)
- [Harness segundo Fowler: 4 camadas no backend](https://kenimoto.dev/pt/blog/fowler-4-harness-invisiveis-backend/)
- [Confiei no Claude: 3 vieses da era IA que peguei](https://kenimoto.dev/pt/blog/confiei-no-claude-3-vieses-nova-era-ia/)
- [Modelos maiores mentem melhor (3.5x specificity)](https://kenimoto.dev/pt/blog/modelos-maiores-mentem-melhor-3-5x-specificity/)
- [PLAID e o dia que Claude Code multiplicou os PRs por 4x — o que a equipe fez para não desabar](https://kenimoto.dev/pt/blog/plaid-claude-code-4x-prs/)
- [Self-Evolving Agent: meu harness reescreveu AGENTS.md 24 vezes em 6 meses](https://kenimoto.dev/pt/blog/self-evolving-agent-24-rewrites-agents-md-6-meses/)
- [Reciprocidade em code review: revisei 47 PRs open-source em 6 meses, meu refactor grande passou em 3 dias](https://kenimoto.dev/pt/blog/reciprocidade-code-review-cialdini-6-meses-oss/)
- [Reunião de decisão técnica que travava 90 min: 3 táticas que a cortaram para 35](https://kenimoto.dev/pt/blog/reuniao-decisao-tecnica-90-para-35-min-3-taticas/)
- [O ataque Shai-Hulud: por que apertar 'Sim' no Claude Code virou a nova falha crítica](https://kenimoto.dev/pt/blog/shai-hulud-claude-code-aprovar-sim-falha-critica/)
- [Agente autônomo por 30 dias: 5 incidentes novos custaram R$ 8.400](https://kenimoto.dev/pt/blog/agente-autonomo-30-dias-5-incidentes-brl/)
- [A 4ª falha de Spec-Driven Development com Claude Code (post-mortem de 6 semanas)](https://kenimoto.dev/pt/blog/4a-falha-spec-driven-development-claude-code-6-semanas/)
- [autoFixable: o padrão que corrige 40% dos comentários de PR antes do humano ver](https://kenimoto.dev/pt/blog/autofixable-padrao-40-porcento-comentarios-pr/)
- [Filtro duplo de IA no PR: CodeRabbit + Claude Code Action cortaram 47% dos comentários humanos (ch10)](https://kenimoto.dev/pt/blog/filtro-duplo-ia-pr-coderabbit-claude-code-action-cortou-47/)
- [Automatizei minha declaração de imposto com Claude Code: 14 horas viraram 47 minutos (e o que a Receita ainda não sabe)](https://kenimoto.dev/pt/blog/automatizei-declaracao-imposto-claude-code-14h-47min/)
- [CTR do #1 do Google caiu 34,5% com AI Overview — o checklist LLMO de 1 hora](https://kenimoto.dev/pt/blog/ai-overview-ctr-34-5-checklist-llmo-1-hora/)
- [A transcrição funcionou. O acompanhamento, não: tentei gerar playback de karaokê com OSS e travei três vezes](https://kenimoto.dev/pt/blog/karaoke-oss-transcricao-funcionou-falta-o-arranjo/)
- [AGENTS.md como política única de code review — 47 PRs depois, sistema virou juiz único](https://kenimoto.dev/pt/blog/agents-md-como-politica-unica-code-review-47-prs/)
- [Agentic RAG em produção: o agente decide 4 vezes na mesma pergunta - alucinação cai 3x, custo dobra](https://kenimoto.dev/pt/blog/agentic-rag-4-iteracoes-3x-menos-alucinacao-2x-custo/)
- [CLAUDE.md pede, hook impõe: os 3 níveis de contrato que separam 90% de 100%](https://kenimoto.dev/pt/blog/claude-md-pede-hook-impoe-90-vs-100/)
- [Troquei Sonnet por Qwen 7B local em 3 tarefas: bill de US$ 47 virou R$ 0 (e 2 tarefas ficaram melhores)](https://kenimoto.dev/pt/blog/troquei-sonnet-por-qwen-7b-local-47-para-0/)
- [Few-shot: testei 2/4/8/16/32 exemplos em 3 tarefas e a qualidade caiu depois de N](https://kenimoto.dev/pt/blog/few-shot-2-4-8-16-32-ponto-otimo-3-tarefas/)
- [Padrão autoFixable: como separei 41% dos comentários de review antes da IA olhar](https://kenimoto.dev/pt/blog/autofixable-antes-ia-41-porcento-comentarios-review-cortados/)
- [Ataque Shai-Hulud no Claude Code: 3 defesas contra o autopiloto 'Yes'](https://kenimoto.dev/pt/blog/ataque-shai-hulud-claude-code-3-defesas-autopiloto-yes/)
- [Haiku 3 + RAG (11,8) venceu Sonnet 4 puro (5,3): o experimento que provou que contexto vale mais que modelo](https://kenimoto.dev/pt/blog/haiku-3-rag-11-8-vs-sonnet-4-puro-5-3-contexto-vence-modelo/)
- [52,8% para 66,5%: como Generator-Evaluator inspirado em GAN eleva accuracy do RAG](https://kenimoto.dev/pt/blog/52-8-para-66-5-generator-evaluator-rag-gan/)
- [Camada 1 → 2 → 3: o funil de code review com IA que corta 70% do trabalho humano](https://kenimoto.dev/pt/blog/camada-1-2-3-funil-code-review-ia-70pct/)
- [O resumo de IA diz que \"não é golpe\". O trampolim é a caixa de busca do seu site](https://kenimoto.dev/pt/blog/resumo-de-ia-nao-e-golpe-caixa-de-busca/)
- [Haiku e Qwen 4B ganharam do Opus em 30% dos meus 47 casos reais](https://kenimoto.dev/pt/blog/haiku-qwen-4b-vencem-opus-14-47/)
- [Dark Patterns em Entrevista Tech: 7 Armadilhas que Identifiquei em 12 Processos](https://kenimoto.dev/pt/blog/sete-dark-patterns-entrevista-tech-12-processos/)
- [Agente 24h e o 2º incidente: 3 falhas que o postmortem inicial escondeu](https://kenimoto.dev/pt/blog/agente-24h-segundo-incidente-3-falhas-postmortem/)
- [3 vieses cognitivos que estragam sua estimativa de sprint — e como o Claude Code corta 2 deles](https://kenimoto.dev/pt/blog/3-vieses-cognitivos-estimativa-sprint-claude-code-corta-2/)
- [TDD com Playwright MCP e CI: pipeline Claude Code em 47 min](https://kenimoto.dev/pt/blog/tdd-playwright-mcp-ci-claude-code-47-min/)
- [Escrevi 'testes antes do PR' no AGENTS.md por 3 Meses. Só 12% Seguiram.](https://kenimoto.dev/pt/blog/agents-md-testes-pr-12-por-cento-seguiram/)
- [GraphRAG para microsserviço: 7 passos para saber quais ADRs quebram seu service](https://kenimoto.dev/pt/blog/graphrag-microsservico-adr-7-passos/)
- [40% dos projetos de agentes de IA falharam em 2026: peguei 12 post-mortems e classifiquei em prompt, context ou harness](https://kenimoto.dev/pt/blog/agentes-ia-40-falharam-2026-12-postmortems-prompt-context-harness/)
- [Trocamos o Harness (Não o Modelo) e o Benchmark Subiu +13,7 Pontos](https://kenimoto.dev/pt/blog/trocamos-harness-nao-modelo-benchmark-137-pontos/)
- [Harness Engineering: os 6 componentes que auditam seu CLAUDE.md em 10 min](https://kenimoto.dev/pt/blog/harness-engineering-6-componentes-auditoria-claude-md/)
- [IA escreve monótono em 70 de 70 células, em 3 línguas. E o corpus humano em português veio do TabNews](https://kenimoto.dev/pt/blog/sotaque-de-maquina-70-70-celulas-3-linguas/)
- [5 mentiras que todo LLM te conta — e como o Context Engineering corrige cada uma](https://kenimoto.dev/pt/blog/5-mentiras-llm-context-engineering-pt/)
- [Prompt morreu, Context está morrendo: Harness Engineering é a próxima onda](https://kenimoto.dev/pt/blog/prompt-morreu-context-morrendo-harness-engineering-proxima-onda/)
- [Harness Engineering: 5 empresas, 5 definições — o mapa dos termos em conflito (2026)](https://kenimoto.dev/pt/blog/harness-engineering-5-empresas-mapa-termos-conflito/)
- [Trabalho híbrido sabota promoção: viés de proximidade (2026)](https://kenimoto.dev/pt/blog/trabalho-hibrido-vies-proximidade-promocao-2026/)
- [Adicionando um servidor MCP a um OSS: do YAML à imagem PNG](https://kenimoto.dev/pt/blog/historymap-mcp-server-yaml-to-png/)
- [Knowledge Graph pessoal: o segundo cérebro que o Notion nunca vai conseguir ser](https://kenimoto.dev/pt/blog/knowledge-graph-pessoal-segundo-cerebro-notion-nunca-vai-conseguir-ser/)
- [Meu agente Claude não percebeu que o cliente estava furioso — a camada de senso comum emocional que Meta descobriu antes de mim](https://kenimoto.dev/pt/blog/agente-claude-cliente-furioso-commonsense-emocional/)
- [Camada 1 do harness de code review: por que 3 hooks pre-commit cortam 40% do trabalho da IA](https://kenimoto.dev/pt/blog/harness-code-review-camada-1-hooks-pre-commit-40-percent/)
- [historymap: um arquivo YAML vira uma linha do tempo estilo site corporativo](https://kenimoto.dev/pt/blog/historymap-yaml-linha-do-tempo-oss/)
- [Padrão autoFixable: 0 revisões humanas de lint em 3 meses](https://kenimoto.dev/pt/blog/padrao-autofixable-zero-revisoes-humanas-lint-3-meses/)
- [README de perfil no GitHub: 7 estilos em 10](https://kenimoto.dev/pt/blog/dissequei-10-perfis-github-7-estilos/)
- [7 passos para construir um Knowledge Graph do seu código e cortar 90% das sugestões fantasma do Cursor](https://kenimoto.dev/pt/blog/7-passos-knowledge-graph-codigo-cursor-alucinacao/)
- [Eu estava lançando meu plugin às 3 da manhã quando ele me avisou: 67% de contexto. Vinte minutos depois, ele mesmo quebrou](https://kenimoto.dev/pt/blog/compact-ops-plugin-me-avisou-no-meio-do-lancamento/)
- [GraphRAG vs RAG clássico: quando o grafo vale 7x o custo](https://kenimoto.dev/pt/blog/graphrag-vs-rag-classico-4-projetos-quando-vale-7x-custo/)
- [Confiei no Claude Code no piloto automático e o Shai-Hulud quase entrou: 5 permissões que reduzi hoje](https://kenimoto.dev/pt/blog/piloto-automatico-5-permissoes/)
- [autoFixable: transformei 30 min de bate-e-volta em revisão de código em 47 segundos](https://kenimoto.dev/pt/blog/autofixable-30min-para-47-segundos-revisao-ia/)
- [RAG vs GraphRAG: o teste de 7 passos](https://kenimoto.dev/pt/blog/graphrag-vs-rag-teste-7-passos-quando-trocar/)
- [Agent = Model + Harness: autópsia de US$ 47.000](https://kenimoto.dev/pt/blog/47k-fim-de-semana-argumento-definitivo-agent-model-harness/)
- [O criador do CLAUDE.md escreve 2 linhas. A comunidade escreve 100. Quem está errado?](https://kenimoto.dev/pt/blog/claude-md-2-vs-100-linhas-quem-erra/)
- [Pare de pagar pelo modelo grande: Haiku + RAG fez 11,8 contra 5,3 do Sonnet sozinho](https://kenimoto.dev/pt/blog/haiku-rag-vence-sonnet-sozinho-contexto/)
- [Codex rodou em 1 milhão de linhas de código: e provou que trocar de modelo nunca conserta seu agente](https://kenimoto.dev/pt/blog/codex-1-milhao-linhas-trocar-modelo-nao-conserta-agente/)
- [Os 6 componentes de um harness de IA: o checklist de auditoria que cabe em uma página](https://kenimoto.dev/pt/blog/6-componentes-harness-checklist-auditoria-uma-pagina/)
- [US$ 47.000 num fim de semana: a tempestade de 2,3 milhões de retries que o harness teria parado em 4 linhas](https://kenimoto.dev/pt/blog/47000-fim-de-semana-tempestade-retries-harness-4-linhas/)
- [Agent = Model + Harness: Por Que Trocar GPT por Claude Nunca Resolve o Bug do Seu Agente](https://kenimoto.dev/pt/blog/agent-equal-model-mais-harness-trocar-modelo-nao-resolve/)
- [MCP me custou 27 mil tokens só pra dizer oi. Knowledge Graph cortou 150 mil pra 18 mil no code review (8,3x). Qual ferramenta ganha qual batalha](https://kenimoto.dev/pt/blog/mcp-27k-vs-kg-8x-qual-ganha/)
- [Construir ficou trivial com Claude Code — o gargalo agora é vender, e quase nenhum engenheiro percebeu](https://kenimoto.dev/pt/blog/construir-ficou-trivial-gargalo-vender/)
- [Pare de mandar o Claude Code codar direto: desenhei o dia no Plan Mode e o retrabalho despencou](https://kenimoto.dev/pt/blog/plan-mode-projetar-dia-retrabalho/)
- [40% dos projetos de agente de IA vão ser cancelados até 2027. A culpa quase nunca é do modelo.](https://kenimoto.dev/pt/blog/agentes-ia-40-falham-harness/)
- [95% do conhecimento da sua empresa mora na cabeça de um veterano: como a Meta transformou o tácito em grafo](https://kenimoto.dev/pt/blog/conhecimento-tacito-grafo-meta/)
- [Anotar conhecimento à mão não escala: montei um pipeline que registrou 300 fontes sozinho em 3 meses](https://kenimoto.dev/pt/blog/banco-conhecimento-300-3-meses/)
- [Seu cérebro reconhece o bug antes de você ler o log (e erra mais do que você imagina)](https://kenimoto.dev/pt/blog/cerebro-reconhece-bug-antes-do-log/)
- [Seu site em React é invisível pra IA: os crawlers não rodam o seu JavaScript](https://kenimoto.dev/pt/blog/site-react-invisivel-ia-crawlers-nao-rodam-javascript/)
- [Deixei a IA escrever 100 testes verdes. Mutation testing diz que pegaram só 58% dos bugs.](https://kenimoto.dev/pt/blog/ia-100-testes-verdes-mutation/)
- [O System Prompt que todo mundo escreve é o que faz a IA mentir](https://kenimoto.dev/pt/blog/responda-em-detalhes-faz-ia-mentir/)
- [796 pacotes npm comprometidos por gente que apertou 'Sim' no automático — a anatomia do Shai-Hulud](https://kenimoto.dev/pt/blog/shai-hulud-796-pacotes-sim-automatico/)
- [Disseram que 'inglês é a nova linguagem de programação'. Passei 3 meses programando assim, e ainda preciso saber programar.](https://kenimoto.dev/pt/blog/ingles-nao-e-a-nova-linguagem-de-programacao/)
- [O cargo de 'engenheiro' não vai sumir — vira 80% decisão, 20% código](https://kenimoto.dev/pt/blog/engenheiro-nao-vai-sumir/)
- [Eu rodava geração de conteúdo e checagem de pedidos no mesmo cron. Uma API lenta travou tudo por 5 horas](https://kenimoto.dev/pt/blog/cron-unico-time-vs-event-5h/)
- [RAG só busca. GraphRAG raciocina. Como o LinkedIn cortou 28,6% do tempo de suporte](https://kenimoto.dev/pt/blog/graphrag-raciocina-linkedin-suporte/)
- [Estimei 47 tarefas em 6 meses: a regra do x2 que eu parei de negar](https://kenimoto.dev/pt/blog/estimativa-x2-47-tarefas/)
- [Uma escritora publicou um app inteiro escrevendo quase zero linha de código: 9.000 linhas, 20 horas](https://kenimoto.dev/pt/blog/escritora-publicou-app-quase-zero-codigo/)
- [Construí tudo no Claude Code. Aí o preço dobrou da noite pro dia](https://kenimoto.dev/pt/blog/dependencia-ferramenta-ia-plano-b/)
- [O passo mais valioso da revisão de código por IA não usa IA nenhuma](https://kenimoto.dev/pt/blog/revisao-codigo-ia-passo-sem-ia-tree-sitter/)
- [Não automatize tudo: os 4 níveis de delegação que decidem se o Claude Code te ajuda ou atrapalha](https://kenimoto.dev/pt/blog/4-niveis-delegacao-claude-code/)
- [Construí o RAG que mandaram fazer (Voyage + vector DB + busca semântica). Depois deletei tudo: o grep venceu.](https://kenimoto.dev/pt/blog/construi-rag-deletei-grep-venceu/)
- [O modelo maior mente melhor: por que o Sonnet inventa com mais confiança que o Haiku](https://kenimoto.dev/pt/blog/modelo-maior-mente-melhor-sonnet-vs-haiku/)
- [Confiei que a IA me deixava mais rápido. Os dados dizem que fiquei mais lento](https://kenimoto.dev/pt/blog/confiei-ia-mais-rapido-fiquei-mais-lento/)
- [O modelo barato + RAG bateu o caro em 223% nos meus testes. Refiz a conta com os preços de 2026.](https://kenimoto.dev/pt/blog/haiku-rag-223-modelo-barato-precos-2026/)
- [Skills do Claude Code consomem tokens mesmo sem disparar. Medi 5 Skills em 7 horas — os 3 que nunca rodaram comeram 11% da conta.](https://kenimoto.dev/pt/blog/skills-3-dormentes-18-tokens/)
- [Coloquei 7 agentes de IA no cron diário. 2 ficaram em silêncio por 18 dias. Tracing não pegou. Um contrato de exit code pegou.](https://kenimoto.dev/pt/blog/cron-7-agentes-silent-18-dias/)
- [Rodei 3 sessões do Claude Code em paralelo por 8 horas. Elas se sobrescreveram 2 vezes. R$ 280 em retrabalho.](https://kenimoto.dev/pt/blog/tres-sessoes-claude-code-paralelo-8h-2-colisoes/)
- [Pedi a 5 IAs que citassem meu próprio blog. 31 artigos publicados, só 3 apareceram.](https://kenimoto.dev/pt/blog/cinco-ias-citaram-meu-blog-tres-de-trinta-e-um/)
- [Coloquei 11 schemas JSON-LD. Em 3 meses, só 3 foram citados por IA](https://kenimoto.dev/pt/blog/11-json-ld-apenas-3-citados/)
- [Por que toda code review vira briga: 5 vieses psicológicos que sabotam seu PR](https://kenimoto.dev/pt/blog/code-review-vira-briga-5-vieses-psicologicos/)
- [Adicionei um 4º agente que audita meus outros agentes. Ele pegou meu Strategist enrolando há 3 semanas.](https://kenimoto.dev/pt/blog/quarto-agente-evolver-pegou-strategist-enrolando/)
- [Quanto um time gasta apontando indentação no PR: a conta dá R$ 3.600 por mês](https://kenimoto.dev/pt/blog/medi-revisao-codigo-3600-reais-mes/)
- [Traduzi o blog pra 4 idiomas. O português pegou quase 4× o tráfego do inglês.](https://kenimoto.dev/pt/blog/traduzi-blog-4-idiomas-portugues-4x-trafego/)
- [Conectei o Claude Code a 4 servidores MCP. Só o handshake queimou 27.000 tokens — e nem foi o pior.](https://kenimoto.dev/pt/blog/conectei-claude-4-servidores-mcp-27k-tokens-handshake/)
- [Contei quantas vezes o Claude me disse 'Você está absolutamente certo!' na semana passada. 47 vezes. Em 11 delas, eu não estava. Nas outras 36, o Claude também não.](https://kenimoto.dev/pt/blog/claude-voce-esta-certo-47-vezes-sycophancy-medi/)
- [Plugei o mesmo site em 7 rastreadores de citação por IA. Nenhum bateu com o outro.](https://kenimoto.dev/pt/blog/7-rastreadores-citacoes-ia-numeros-diferentes/)
- [Os 5 crawlers de IA que mais bateram nos meus sites em 30 dias - o que os logs revelaram sobre LLMO](https://kenimoto.dev/pt/blog/cinco-crawlers-ia-bateram-meu-site-30-dias/)
- [Pluguei o Claude num MCP de Chaos. Matou staging 4 vezes pra achar um bug que ignorávamos há 6 meses.](https://kenimoto.dev/pt/blog/claude-chaos-engineering-mcp-matou-staging-4-vezes/)
- [Peguei o Claude escondendo meu bug 3 vezes seguidas. Aí virei 10 hábitos de debug em prompts.](https://kenimoto.dev/pt/blog/claude-escondeu-meu-bug-3-vezes-10-habitos-debug/)
- [Dei WebSearch ao meu Strategist. 5 temas levaram 20 minutos. Separar em 3 agentes baixou pra 3.](https://kenimoto.dev/pt/blog/tres-papeis-observer-strategist-marketer-separacao/)
- [Testei 5 stacks de Voice AI. Apenas 2 ficaram abaixo de 300ms.](https://kenimoto.dev/pt/blog/cinco-stacks-voice-ai-apenas-dois-abaixo-300ms/)
- [Coloquei 3 sub-agentes do Claude Code para revisar o mesmo PR. Discordaram em 41% dos comentários.](https://kenimoto.dev/pt/blog/tres-sub-agentes-revisaram-mesmo-pr-40-discordancia/)
- [Auditei 30 arquivos llms.txt em produção. 5 anti-padrões já estão se formando.](https://kenimoto.dev/pt/blog/auditei-30-arquivos-llms-txt-5-anti-padroes/)
- [Eu troquei Claude Code por OpenClaw por 24 horas. Veja o que quebrou — e o que melhorou](https://kenimoto.dev/pt/blog/openclaw-claude-code-24h-troca/)
- [Eu me recusei a escrever spec até o Claude gerar o código errado 3 vezes](https://kenimoto.dev/pt/blog/spec-driven-development-claude-code-3-falhas/)
- [Deixei meu agente Claude Code rodar 24 horas sozinho. A conta de R$ 2.000 nem foi o pior.](https://kenimoto.dev/pt/blog/agente-ia-24-horas-incidentes-seguranca/)
- [Por que o ChatGPT ignora o seu site (mesmo se você for #1 no Google)](https://kenimoto.dev/pt/blog/chatgpt-ignora-seu-site-llmo/)
- [Por que troquei prompt engineering por context engineering](https://kenimoto.dev/pt/blog/contexto-vs-prompt-engineering/)
- [Cursor vs Claude Code: 6 meses e a volta](https://kenimoto.dev/pt/blog/parei-cursor-voltei-terminal/)

### Spanish (Español)

- [Tree-sitter code review: 3 pasos sin IA](https://kenimoto.dev/es/blog/tree-sitter-code-review-3-pasos-sin-ia/)
- [Claude Code: 3 patrones de memoria post-compact](https://kenimoto.dev/es/blog/claude-code-3-patrones-memoria-post-compact/)
- [Servidor MCP silencioso: 3 patrones de debug](https://kenimoto.dev/es/blog/mcp-servidor-silencioso-3-patrones-debug/)
- [Code review IA en 3 capas: hooks + IA + humano](https://kenimoto.dev/es/blog/code-review-ia-3-capas-hooks-ia-humano/)
- [OpenAI Realtime API: bajar tu voz IA a 300ms](https://kenimoto.dev/es/blog/openai-realtime-api-bajar-voz-ia-300ms/)
- [nanochat CORE: 22 Tests en 1 Número (0.256525)](https://kenimoto.dev/es/blog/nanochat-core-score-22-tests-un-numero/)
- [GraphRAG explicado: 7 pasos, vector vs multi-hop](https://kenimoto.dev/es/blog/graphrag-7-pasos-multi-hop/)
- [AGENTS.md vs CLAUDE.md en 2026: árbol de decisión de 5 preguntas](https://kenimoto.dev/es/blog/agents-md-vs-claude-md-2026-arbol-decision-5-preguntas/)
- [MCP en 5 minutos para devs LatAm: 1 protocolo, 12 integraciones](https://kenimoto.dev/es/blog/mcp-en-5-minutos-devs-latam-1-protocolo-12-integraciones/)
- [Claude Code Hooks v2 en LatAm: 5 usos y 30+ eventos](https://kenimoto.dev/es/blog/claude-code-hooks-v2-5-usos-produccion-latam/)
- [3 sesgos que rompen tus estimaciones cuando la IA dice \"lo hago en 10 minutos\"](https://kenimoto.dev/es/blog/3-sesgos-estimacion-ia-dice-10-minutos/)
- [Code review con Knowledge Graph + MCP: pasé de 150k a 18k tokens (8x menos que RAG del repo)](https://kenimoto.dev/es/blog/code-review-graph-mcp-8x-menos-tokens/)
- [El umbral de 300ms en IA de voz: los 3 límites de Nielsen aplicados al audio](https://kenimoto.dev/es/blog/300ms-ia-voz-3-limites-nielsen-audio/)
- [Contexto pequeño vence a modelo gigante: 3 experimentos en 2026](https://kenimoto.dev/es/blog/contexto-pequeno-vence-modelo-gigante-3-experimentos/)
- [Cron para agentes de IA: 4 patrones para detectar fallos silenciosos](https://kenimoto.dev/es/blog/cron-agentes-ia-4-patrones-fallos-silenciosos/)
- [Codex CLI vs Claude Code: probé 7 tareas reales en el mismo repo (guía LatAm 2026-08)](https://kenimoto.dev/es/blog/codex-cli-vs-claude-code-7-tareas-reales-latam-2026/)
- [Detectar 'AI Slop' en tu UI: 7 frases que delatan al LLM en tu copy](https://kenimoto.dev/es/blog/detectar-ai-slop-ui-7-frases-llm-copy/)
- [Bias de automatización — por qué confías en Claude más que en tu colega senior](https://kenimoto.dev/es/blog/bias-automatizacion-confias-claude-mas-colega-senior/)
- [Claude Code Skills vs MCP servers: cuándo usar cada uno (guía práctica LatAm 2026)](https://kenimoto.dev/es/blog/claude-code-skills-vs-mcp-servers-guia-latam-2026/)
- [Cuando tu agente de IA se atasca en un bucle: 4 síntomas y el fix para cada uno](https://kenimoto.dev/es/blog/agente-ia-bucle-4-sintomas-fix/)
- [Few-shot en LLMs: cuántos ejemplos son el punto óptimo antes de que la calidad caiga (por tarea)](https://kenimoto.dev/es/blog/few-shot-punto-optimo-cuantos-ejemplos-por-tarea/)
- [GraphRAG vs RAG clásico: 3 métricas que cambian cuando el grafo entra en juego](https://kenimoto.dev/es/blog/graphrag-vs-rag-clasico-3-metricas-grafo-entra/)
- [Medí las 346 réplicas de un M7.1 en Japón. Después lo comparé con el 19S y con Chile.](https://kenimoto.dev/es/blog/346-replicas-japon-19s-chile-misma-ley/)
- [GraphRAG vs RAG tradicional: cuándo el costo extra vale (y cuándo no)](https://kenimoto.dev/es/blog/graphrag-vs-rag-tradicional-cuando-el-costo-extra-vale/)
- [Entrenar un GPT-2 costaba $43,000 en 2019: cuánto cuesta hoy con nanochat de Karpathy (precios GPU 2026)](https://kenimoto.dev/es/blog/nanochat-entrenar-gpt-2-costaba-43000-2019-cuanto-cuesta-hoy/)
- [Auditar 30 blogs de devs con llmoframework: 4 de 5 mejores fallan las mismas 3 comprobaciones](https://kenimoto.dev/es/blog/auditar-30-blogs-devs-llmoframework-4-de-5-fallan-mismos-3/)
- [Prompt, Contexto, Harness: las 3 etapas de evolución que casi nadie termina](https://kenimoto.dev/es/blog/prompt-contexto-harness-3-etapas-evolucion/)
- [Qwen 35B en una RTX 4070: 2 Flags para Pasar de 12 a 34 tok/s (2.8x)](https://kenimoto.dev/es/blog/qwen-35b-rtx-4070-dos-flags-2-8x/)
- [GraphRAG vs RAG clásico: guía práctica para elegir en 5 preguntas](https://kenimoto.dev/es/blog/graphrag-vs-rag-5-preguntas-decidir/)
- [Sponsorware desde LatAm: 90 días hasta $1000/mes en GitHub Sponsors partiendo de cero](https://kenimoto.dev/es/blog/sponsorware-latam-90-dias-1000-mensuales-github-sponsors/)
- [Cuantización KV cache: 8x más contexto en RTX 4070 con dos flags](https://kenimoto.dev/es/blog/cuantizacion-kv-cache-8x-contexto-rtx-4070/)
- [System 1 vs System 2 al Depurar: Bugs de Madrugada Cuestan 3× Más](https://kenimoto.dev/es/blog/system-1-vs-system-2-debugging-madrugada/)
- [Nanochat de Karpathy: cuánto cuesta hoy entrenar un modelo tipo GPT-2 (era $43,000 en 2019)](https://kenimoto.dev/es/blog/nanochat-karpathy-costo-entrenar-gpt2-2019-2026/)
- [GraphRAG vs RAG: 7 pasos para tu primer knowledge graph de código en 30 min](https://kenimoto.dev/es/blog/graphrag-vs-rag-7-pasos-knowledge-graph-codigo/)
- [Prompt Caching de Anthropic: el 90% de descuento y 3 casos donde NO conviene](https://kenimoto.dev/es/blog/prompt-caching-anthropic-es/)
- [MCP + USB Serial: Claude Code controla hardware (2026)](https://kenimoto.dev/es/blog/mcp-usb-serial-claude-code-microcontrolador/)
- [Harness Engineering: la tercera era después de Prompt y Context, en 3 señales](https://kenimoto.dev/es/blog/harness-engineering-tercera-era-prompt-context-3-senales/)
- [Grafo de conocimiento personal en Neo4j: 300 notas sin mensualidad](https://kenimoto.dev/es/blog/grafo-conocimiento-personal-neo4j-notion/)
- [Los 7 archivos que Claude Code lee antes de escribir código — checklist de context engineering en 5 minutos](https://kenimoto.dev/es/blog/7-archivos-claude-code-context-engineering-checklist-5-minutos/)
- [Cuando el agente borró producción y cómo el harness lo detuvo en 3 capas](https://kenimoto.dev/es/blog/agente-borro-produccion-harness-3-capas/)
- [historymap: un archivo YAML se convierte en una línea de tiempo estilo sitio corporativo](https://kenimoto.dev/es/blog/historymap-yaml-linea-de-tiempo-oss/)
- [Dejé un agente de IA corriendo 24 horas: 3 incidentes de seguridad y los 7 permisos que hoy bloqueo antes de arrancar](https://kenimoto.dev/es/blog/agente-ia-24h-3-incidentes-7-permisos-bloqueo/)
- [Fatiga de decisiones con 3 IDEs de IA abiertos: el experimento de Vohs traducido a nuestro trabajo](https://kenimoto.dev/es/blog/fatiga-decisiones-3-ides-ia-vohs-traducido-a-dev/)
- [Agente = Modelo + Harness: la matriz de 6 decisiones para elegir donde invertir tu tiempo](https://kenimoto.dev/es/blog/agente-modelo-harness-matriz-6-decisiones/)
- [Construí mi base de conocimiento personal: 300 fuentes en 3 meses (lo que sirvió y lo que perdió tiempo)](https://kenimoto.dev/es/blog/300-fuentes-3-meses-base-conocimiento-personal/)
- [RAG vs GraphRAG: matriz de 4 preguntas](https://kenimoto.dev/es/blog/rag-vs-graphrag-matriz-decision/)
- [Los 6 componentes de un harness de agente IA: audita tu CLAUDE.md en 10 minutos](https://kenimoto.dev/es/blog/6-componentes-harness-auditar-claude-md-10-minutos/)
- [Modelo pequeño + buen contexto le gana al modelo grande solo (y cuesta 70% menos)](https://kenimoto.dev/es/blog/modelo-pequeno-buen-contexto-gana-70-menos-costo/)
- [Las 5 preguntas que separan Prompt Engineering de Context Engineering (con ejemplos en Claude Code)](https://kenimoto.dev/es/blog/5-preguntas-context-vs-prompt-engineering-claude-code/)
- [Context Engineering vs Prompt Engineering: Por Qué Reescribir el Prompt No Sirve y Qué Sí (Benchmark 4.6x)](https://kenimoto.dev/es/blog/context-engineering-vs-prompt-engineering-benchmark-4-6x/)
- [MCP no soporta subir archivos: probé 7 servidores y cada uno se inventó un workaround. 5 de los 7 abren un hueco de seguridad medible](https://kenimoto.dev/es/blog/mcp-no-soporta-archivos-7-servidores/)
- [Plan Mode de Claude Code: aprueba el plan antes de ejecutar (guía práctica)](https://kenimoto.dev/es/blog/plan-mode-aprobar-antes-de-ejecutar/)
- [Simular una red mala en tus pruebas E2E de WebRTC: pérdida de paquetes con tc/netem, sin infraestructura cara](https://kenimoto.dev/es/blog/simular-red-mala-webrtc-tc-netem/)
- [Tu backend dice p99=50ms y el usuario ve 3 segundos: la asimetría que solo el trace_id resuelve](https://kenimoto.dev/es/blog/asimetria-observabilidad-trace-id/)
- [Por qué tu sistema multi-agente no escala: la arquitectura de memoria decide el techo (25 agentes contra 1.000.000)](https://kenimoto.dev/es/blog/memoria-decide-techo-multiagente/)
- [El modelo no cambió, pero la precisión subió de 52.8% a 66.5%: qué es la ingeniería de harness y por qué invertir en el andamiaje le gana a cambiar de modelo](https://kenimoto.dev/es/blog/harness-52-a-66-mismo-modelo/)
- [Cómo darle memoria persistente a tu agente de IA: la arquitectura de 7 archivos (y por qué cargar todo es un error)](https://kenimoto.dev/es/blog/memoria-persistente-agente-7-archivos/)
- [Por qué los crawlers de IA no ven tu sitio en JavaScript (y cómo arreglarlo con SSR o prerender)](https://kenimoto.dev/es/blog/crawlers-ia-no-ven-javascript-ssr-prerender/)
- [Le escribí tests a mi LLMO: verificar con Playwright que los crawlers de IA leen cada página](https://kenimoto.dev/es/blog/tests-llmo-playwright-crawlers/)
- [‘Casi siempre’ y ‘siempre sin excepción’ no son lo mismo: cómo hacer que tu code review con IA corra al 100%](https://kenimoto.dev/es/blog/code-review-ia-al-100-con-hooks/)
- [LLMO en 15 minutos: llms.txt + JSON-LD para que la IA pueda encontrar tu sitio](https://kenimoto.dev/es/blog/llmo-15-minutos-llms-txt-json-ld/)
- [Por qué tu agente de IA no tiene 'sentido común' (y cómo se lo das con un knowledge graph)](https://kenimoto.dev/es/blog/sentido-comun-agente-ia-knowledge-graph/)
- [Anatomía de la latencia en agentes de voz: dónde se te van los 300ms (STT → LLM → TTS)](https://kenimoto.dev/es/blog/anatomia-latencia-voz-300ms/)
- [Etiquetar nit vs must en cada PR: cómo bajé el merge de 48h a 24h](https://kenimoto.dev/es/blog/conventional-comments-nit-must-48h-24h/)
- [Brave Search API para agentes de IA en 15 min](https://kenimoto.dev/es/blog/brave-search-llm-context-api-agente-ia/)
- [Corrí un LLM local en mi GPU sin pagar API: la regla VRAM-a-modelo que evita el OOM](https://kenimoto.dev/es/blog/llm-local-gpu-vram-modelo/)
- [Few-shot no le enseña conocimiento a tu LLM: le enseña a decir \"no lo sé\"](https://kenimoto.dev/es/blog/few-shot-no-ensena-conocimiento/)
- [Cómo blindar los permisos de Claude Code para que tu agente no filtre secretos](https://kenimoto.dev/es/blog/permisos-claude-code-evitar-fuga-secretos/)
- [Por qué Claude ignora tu CLAUDE.md 1 de cada 20 veces (y cómo lo arreglé con exit code 2)](https://kenimoto.dev/es/blog/claude-code-hooks-exit-code-2-reglas-deterministas/)
- [Dejé de agregar contexto a mi agente: podar las salidas de herramientas recuperó la precisión](https://kenimoto.dev/es/blog/dejar-de-agregar-contexto-podar-precision/)
- [\"99,5% de disponibilidad\" y \"5.000 pagos fallidos\" son el mismo hecho: el framing del reporte cambia la urgencia](https://kenimoto.dev/es/blog/framing-reporte-incidente-99-5/)
- [Code review a las 3pm: aprobación casi 0%](https://kenimoto.dev/es/blog/fatiga-decision-3pm-code-review-aprobacion/)
- [El ranking de tu página no le importa a la IA — solo cita pasajes](https://kenimoto.dev/es/blog/ranking-pagina-no-importa-ia-cita-pasajes/)
- [Context rot en agentes: 7 pasos para mantener limpia la ventana de contexto (empieza antes de lo que crees)](https://kenimoto.dev/es/blog/context-rot-agentes-7-pasos-ventana-limpia/)
- [Los Claude Code Skills consumen tokens aunque no se activen. Medí 5 Skills durante 7 horas — los 3 que nunca dispararon se llevaron el 11%.](https://kenimoto.dev/es/blog/skills-3-dormidos-18-tokens/)
- [Programé 7 agentes de IA con cron diario. 2 fallaron en silencio durante 18 días. El tracing no los detectó. Un contrato de exit code sí.](https://kenimoto.dev/es/blog/cron-7-agentes-18d-silencioso/)
- [Ejecuté 3 sesiones de Claude Code en paralelo durante 8 horas. Se sobrescribieron 2 veces. Guía LatAm de orquestación segura.](https://kenimoto.dev/es/blog/tres-sesiones-claude-code-paralelo-8h-2-colisiones/)
- [Le pedí a 5 IAs que citaran mi propio blog. 31 artículos publicados, solo 3 aparecieron.](https://kenimoto.dev/es/blog/cinco-ias-citaron-mi-blog-tres-de-treinta-y-uno/)
- [Agregué 11 schemas JSON-LD a mi blog. Tres meses después, solo 3 aparecieron en las citas de IA.](https://kenimoto.dev/es/blog/11-json-ld-solo-3-citados/)
- [Le pedí a Claude que refactorizara 100 funciones. 7 quedaron más lentas en producción](https://kenimoto.dev/es/blog/claude-refactor-100-funciones-7-mas-lentas-produccion/)
- [Le dije a Claude Code que hiciera TDD. Escribió el test DESPUÉS del código 6 de 10 veces](https://kenimoto.dev/es/blog/claude-code-tdd-test-despues-codigo-6-de-10/)
- [Agregué un 4to agente que audita a mis otros agentes. Detectó que mi Strategist llevaba 3 semanas postergando.](https://kenimoto.dev/es/blog/cuarto-agente-evolver-detecto-strategist-postergando/)
- [Traduje mi blog a 4 idiomas. El portugués recibió casi 4× más tráfico que el inglés.](https://kenimoto.dev/es/blog/traduje-blog-4-idiomas-portugues-4x-trafico/)
- [¿Cuánto cuesta realmente un agente de IA al mes? API vs suscripción vs local — el punto de equilibrio en USD](https://kenimoto.dev/es/blog/cuanto-cuesta-agente-ia-al-mes-api-suscripcion-local-punto-equilibrio/)
- [Conté cuántas veces Claude me dijo '¡Tienes toda la razón!' la semana pasada. 47 veces. En 11 de ellas, yo no la tenía. En las otras 36, Claude tampoco.](https://kenimoto.dev/es/blog/claude-tienes-razon-47-veces-sycophancy-medi/)
- [Conecté el mismo sitio a 7 rastreadores de citas de IA. Ninguno coincidió con otro.](https://kenimoto.dev/es/blog/7-rastreadores-citas-ia-numeros-diferentes/)
- [Los 5 crawlers de IA que más golpearon mis sitios en 30 días - lo que los logs me dijeron sobre LLMO](https://kenimoto.dev/es/blog/cinco-crawlers-ia-golpearon-mi-sitio-30-dias/)
- [Conecté Claude a un MCP server de Chaos Engineering. Mató el staging 4 veces antes de encontrar el bug que llevábamos 6 meses ignorando.](https://kenimoto.dev/es/blog/claude-chaos-engineering-mcp-mato-staging-4-veces/)
- [Pillé a Claude escondiendo mi bug 3 veces seguidas. Después convertí 10 hábitos de debug en prompts.](https://kenimoto.dev/es/blog/claude-escondio-mi-bug-3-veces-10-habitos-debug/)
- [Dividí mi agente en 3 roles y elegir 5 temas pasó de 20 a 3 minutos](https://kenimoto.dev/es/blog/tres-roles-observer-strategist-marketer-separacion/)
- [Medí 5 stacks de Voice AI. Solo 2 se mantuvieron bajo los 300ms.](https://kenimoto.dev/es/blog/cinco-stacks-voice-ai-solo-dos-bajo-300ms/)
- [Pedí a 3 sub-agentes de Claude Code que revisaran el mismo PR. Estuvieron en desacuerdo en el 41% de los comentarios.](https://kenimoto.dev/es/blog/tres-sub-agentes-revisaron-mismo-pr-40-desacuerdo/)
- [Audité 30 archivos llms.txt en producción. 5 anti-patrones ya se están formando.](https://kenimoto.dev/es/blog/auditoria-30-archivos-llms-txt-5-anti-patrones/)
- [OpenClaw, Claude Code y Cursor: la guía LatAm para elegir tu primer agente autónomo en 2026](https://kenimoto.dev/es/blog/openclaw-claude-code-cursor-guia-latam/)
- [Spec-Driven Development con asistentes de IA: la guía LatAm para escribir specs antes del primer prompt](https://kenimoto.dev/es/blog/spec-driven-development-asistentes-ia-guia-latam/)
- [Dejé a mi agente Claude Code corriendo 24 horas. La cuenta de USD 400 fue lo de menos.](https://kenimoto.dev/es/blog/agente-ia-autonomo-24-horas-seguridad/)
- [Por qué dejé el prompt engineering por la ingeniería de contexto](https://kenimoto.dev/es/blog/ingenieria-de-contexto-vs-prompt/)

## Research Papers (Zenodo)

- [Rhythmic Monotonization in LLM-Generated Text Is Cross-Lingual — DOI:10.5281/zenodo.21424903](https://doi.org/10.5281/zenodo.21424903)
- [Lexical and Rhythmic Fingerprints of Japanese LLM-Generated Text — DOI:10.5281/zenodo.21413035](https://doi.org/10.5281/zenodo.21413035)
- [When Free Executors Cost More: The Free-Executor Paradox in Iterative LLM Code-Repair Loops — DOI:10.5281/zenodo.20978074](https://doi.org/10.5281/zenodo.20978074)
- [Excess Vocabulary in Japanese AI-Generated Text — DOI:10.5281/zenodo.19233934](https://doi.org/10.5281/zenodo.19233934)
- [AI Text Slop: Stylistic Convergence Across Six LLMs — DOI:10.5281/zenodo.19173035](https://doi.org/10.5281/zenodo.19173035)
- [AI Blue: Color Recognition Bias in Vision-Language Models — DOI:10.5281/zenodo.19159702](https://doi.org/10.5281/zenodo.19159702)

## Citation Preferred

> When citing this site, prefer these canonical entry points per topic.

### Featured Articles

- [The Shai-Hulud npm attack: why signature verification, npm audit, and --ignore-scripts all failed](https://kenimoto.dev/blog/shai-hulud-npm-supply-chain-why-verification-failed/)
- [npmサプライチェーン攻撃Shai-Hulud：署名検証もSBOMも無意味だった理由](https://kenimoto.dev/ja/blog/shai-hulud-npm-supply-chain-why-verification-failed/)
- [Claude Code: The Operator's Field Guide](https://kenimoto.dev/blog/claude-code/)
- [Claude Code 実戦運用ガイド ― エージェントを2年回して分かった8つの勘所](https://kenimoto.dev/ja/blog/claude-code/)
- [LLMO 総合ガイド ― AI 検索に引用される技術を、30ヶ月分の計測から組み立てる](https://kenimoto.dev/ja/blog/llmo/)
- [Por que o ChatGPT ignora o seu site (mesmo se você for #1 no Google)](https://kenimoto.dev/pt/blog/chatgpt-ignora-seu-site-llmo/)
- [Por que troquei prompt engineering por context engineering](https://kenimoto.dev/pt/blog/contexto-vs-prompt-engineering/)
- [Cursor vs Claude Code: 6 meses e a volta](https://kenimoto.dev/pt/blog/parei-cursor-voltei-terminal/)
- [Por qué dejé el prompt engineering por la ingeniería de contexto](https://kenimoto.dev/es/blog/ingenieria-de-contexto-vs-prompt/)
- [9 Bugs in My AI Pipeline: None Were the AI's Fault](https://kenimoto.dev/blog/9-bugs-in-my-ai-pipeline/)
- [AIパイプラインが壊れる9つの理由 -- 全部AIの外側だった](https://kenimoto.dev/ja/blog/9-bugs-in-my-ai-pipeline/)
- [Building an Autonomous Content Pipeline with Claude Code](https://kenimoto.dev/blog/hello-world/)
- [Claude Codeで自律型コンテンツパイプラインを構築した話](https://kenimoto.dev/ja/blog/hello-world/)

### Primary Book LPs (English)

- [Chaos Engineering in Practice](https://kenimoto.dev/books/chaos-engineering-guide/)
- [Practical Claude Code](https://kenimoto.dev/books/claude-code-mastery/)
- [Turning LLMs from Liars into Experts](https://kenimoto.dev/books/context-engineering/)
- [Debugging Engineering](https://kenimoto.dev/books/debugging-engineering/)
- [Why Do Some Words Stay With You Forever?](https://kenimoto.dev/books/engineer-it-quotes/)
- [Cognitive Bias for Software Engineers](https://kenimoto.dev/books/engineer-psychology-tricks/)
- [Systematizing AI Code Review](https://kenimoto.dev/books/harness-code-review/)
- [Harness Engineering](https://kenimoto.dev/books/harness-engineering-guide/)

### Primary Book LPs (日本語)

- [なぜAI生成UIは全部青いのか](https://kenimoto.dev/ja/books/ai-slop-escape-guide/)
- [AIくさい文章から脱出する技術](https://kenimoto.dev/ja/books/ai-text-slop-escape/)
- [カオスエンジニアリング実践ガイド](https://kenimoto.dev/ja/books/chaos-engineering-guide/)
- [Claude Code で動かすマイコン](https://kenimoto.dev/ja/books/claude-code-hardware-control/)
- [実践Claude Code](https://kenimoto.dev/ja/books/claude-code-mastery/)
- [Claude Codeクイックスタート](https://kenimoto.dev/ja/books/claude-code-quickstart/)
- [LLMを「嘘つき」から「専門家」に変える技術](https://kenimoto.dev/ja/books/context-engineering/)
- [デバッグの技法](https://kenimoto.dev/ja/books/debugging-engineering/)

## Side Projects

- [legacydram](https://legacydram.com/) — A whisky curation media reading every bottle as somebody's commit history.
- [ainativemeo](https://ainativemeo.com/) — AI-Native MEO. Local-business optimization for the era when customers ask AI, not search engines, for recommendations. Bilingual EN / JA.

## Weekly Ship Products

> Small apps, tools and games shipped weekly under `/products/<slug>/`.
> Each entry keeps its build log; status shows whether it is live, static (frozen), archived, or promoted to its own domain.
> Catalog: [/products/](https://kenimoto.dev/products/)

### Live

- [hook-chain-lens](https://kenimoto.dev/products/hook-chain-lens/) — A read-only CLI that reads every Claude Code hook scope (user, project, local, plugin) and prints what will actually fire in this cwd, in what order, from where.
- [private-lint](https://kenimoto.dev/products/private-lint/) — Git-hook gate that blocks personal and private names (family names, personal emails, client domains) before commit and push. Pure Bash; detection patterns stay machine-local, never in any repo, and a full-history audit covers the moment before you flip a repository public.
- [quake-lens](https://kenimoto.dev/products/quake-lens/) — Earthquake statistics CLI + MCP server in pure-stdlib Python: Gutenberg-Richter b-value (Aki MLE) and Omori-Utsu aftershock decay (Ogata MLE) from public catalogs.
  - kenimoto-dev: https://kenimoto.dev/ja/blog/uto-m71-yoshin-346-fukuoka-34-jissoku/
  - kenimoto-dev: https://kenimoto.dev/ja/blog/jishin-yochi-vs-yoshin-yosoku-jissoku/
  - kenimoto-dev: https://kenimoto.dev/blog/uto-m71-aftershock-five-days-fukuoka/
  - kenimoto-dev: https://kenimoto.dev/blog/earthquake-prediction-vs-aftershock-forecasting/
  - kenimoto-dev: https://kenimoto.dev/es/blog/346-replicas-japon-19s-chile-misma-ley/
- [rhythm-lens](https://kenimoto.dev/products/rhythm-lens/) — Measures the rhythm of Japanese, English and Portuguese Markdown (sentence-length burstiness, paragraph structure) against measured human/AI distributions from two published papers. A writing feedback instrument, not an AI detector.
  - kenimoto-dev: https://kenimoto.dev/ja/blog/machine-accent-3-languages-70-70-cells/
  - kenimoto-dev: https://kenimoto.dev/blog/machine-accent-3-languages-70-70-cells/
  - kenimoto-dev: https://kenimoto.dev/pt/blog/sotaque-de-maquina-70-70-celulas-3-linguas/
  - kenimoto-dev: https://kenimoto.dev/ja/blog/rhythm-lens-cli-own-article-2-warnings/
  - kenimoto-dev: https://kenimoto.dev/ja/blog/japanese-ai-smell-vocabulary-vs-rhythm-fingerprints/
- [claude-shift](https://kenimoto.dev/products/claude-shift/) — Multi-account Claude Code manager. Switch the active login, sync the two-file auth Claude Code actually reads (~/.claude/.credentials.json + ~/.claude.json), and watch 5-hour and weekly usage across all accounts. CLI + local API + Chrome extension + browser Web UI + Tauri v2 desktop.
  - kenimoto-dev: https://kenimoto.dev/ja/blog/tauri-v2-gawa-wsl-mirrored-networking/
- [opencut-mcp](https://kenimoto.dev/products/opencut-mcp/) — MCP server that drives OpenCut classic from any MCP client. Playwright holds the editor session; twelve tools expose the timeline, media assets, and export pipeline. Runs against a fork of OpenCut classic.
  - kenimoto-dev: /blog/opencut-classic-mcp-4-traps-editor-core-fork/
  - kenimoto-dev: /ja/blog/opencut-classic-mcp-4-traps-editor-core-fork/
  - zenn: https://zenn.dev/kenimo49/articles/opencut-classic-mcp-4-traps
- [mcp-scorecard](https://kenimoto.dev/products/mcp-scorecard/) — Pre-flight checks for MCP servers: passive token footprint, use-case scoping, security, and name safety. LLM-facing quality scorecard.
  - kenimoto-dev: https://kenimoto.dev/blog/pre-flight-your-mcp-4-layers-scorecard/
  - kenimoto-dev: https://kenimoto.dev/ja/blog/pre-flight-your-mcp-4-layers-scorecard/
- [404 Games](https://kenimoto.dev/notexist?g404=menu) — Ten single-file canvas mini games for your 404 page, inspired by famous error pages. Vanilla JS, zero dependencies, theme-aware. Live on this site's 404.
  - zenn: https://zenn.dev/kenimo49/articles/repo-health-10-dimensions-two-fs
- [historymap](https://kenimoto.dev/products/historymap/) — Turn a YAML file into a corporate-style product-history timeline. Ten layouts, switchable right on the page. Static, self-contained, iframe-embeddable.
  - kenimoto-dev: /blog/historymap-yaml-corporate-timeline-oss/
  - kenimoto-dev: /ja/blog/historymap-yaml-timeline-oss/
  - kenimoto-dev: /pt/blog/historymap-yaml-linha-do-tempo-oss/
  - kenimoto-dev: /es/blog/historymap-yaml-linea-de-tiempo-oss/
  - qiita: https://qiita.com/kenimo49/items/159f75910f900f3e71f1
  - kenimoto-dev: /ja/blog/historymap-mcp-server-yaml-to-png/
  - kenimoto-dev: /pt/blog/historymap-mcp-server-yaml-to-png/
- [rag-retriever-bench](https://kenimoto.dev/products/rag-retriever-bench/) — Benchmark harness for RAG retrieval backends: same corpus, same queries, same metrics. Swap the database and see which one fits your workload, with measured numbers instead of vendor benchmarks.
  - qiita: https://qiita.com/kenimo49/items/c5c64b836eabc5828948
- [rag-db-advisor](https://kenimoto.dev/products/rag-db-advisor/) — A RAG that answers RAG-stack questions. Ask which vector backend fits your workload and get measured evidence from rag-retriever-bench, served over MCP or CLI.
  - qiita: https://qiita.com/kenimo49/items/c5c64b836eabc5828948
- [compact-ops](https://kenimoto.dev/products/compact-ops/) — A transparent Claude Code plugin that keeps sessions intact across context compaction. It writes a structured state file before every /compact and injects recovery guidance right after compaction and on resume.
  - devto: https://dev.to/kenimo49/the-compaction-plugin-i-was-releasing-warned-me-mid-release-2f25
  - kenimoto-dev: /ja/blog/compact-ops-dogfooding-67-percent-warn/
  - kenimoto-dev: /pt/blog/compact-ops-plugin-me-avisou-no-meio-do-lancamento/
- [speech-habit-lens](https://kenimoto.dev/products/speech-habit-lens/) — See the habits in a 1-minute speech through three-layer analysis: AmiVoice ESAS 20 emotion parameters, recognized text, and the cross-layer correlations between voice and words. CLI + Streamlit UI.
  - zenn: https://zenn.dev/kenimo49/articles/amivoice-esas-claude-self-speech-analysis
  - zenn: https://zenn.dev/kenimo49/articles/amivoice-esas-claude-4-pitch-comparison
- [domain-pre-flight](https://kenimoto.dev/products/domain-pre-flight/) — Pre-flight checks before registering a domain: structure, TLD risk, past-content history, typosquat similarity, and an aggregate go/no-go verdict. Python CLI + MCP server.
  - zenn: https://zenn.dev/kenimo49/articles/cli-mcp-server-subprocess-vs-library

## Open Source

- [domain-pre-flight](https://github.com/kenimo49/domain-pre-flight) — Python CLI · MIT. A pre-flight check before you register a domain for a new site or app. Structure, history, typosquat, multi-language semantics, LLMO, and trademark deeplinks in one command.
- [voice-clone](https://github.com/kenimo49/voice-clone) — Python · MIT. A voice-cloning tool built on Qwen3-TTS. Clone a voice from ~3 seconds of audio and have any text read in that voice. WSLg recording, multi-language.
- [speech-habit-lens](https://github.com/kenimo49/speech-habit-lens) — Python CLI · MIT. A 1-minute speech habit analyzer combining AmiVoice ESAS (20 acoustic emotion parameters) with an LLM. Three layers — acoustic, textual, and the cross-layer that links body and language — rendered as a Markdown report.
- [persona-hub](https://github.com/kenimo49/persona-hub) — TypeScript + Python · Apache-2.0. A two-part persona hub: a lightweight TypeScript SDK that scores quiz answers locally, and an optional FastAPI service that persists results so the same user can carry their profile across services. Pre-alpha.

## Content URLs

- [EN Profile (canonical entry point)](https://kenimoto.dev/)
- [JA Profile (日本語プロフィール起点)](https://kenimoto.dev/ja/)
- [JA About (日本語紹介ページ)](https://kenimoto.dev/ja/about/)
- [EN Blog](https://kenimoto.dev/blog/)
- [JA Blog](https://kenimoto.dev/ja/blog/)
- [PT Blog](https://kenimoto.dev/pt/blog/)
- [ES Blog](https://kenimoto.dev/es/blog/)
- [RSS (EN)](https://kenimoto.dev/blog.xml)
- [RSS (JA)](https://kenimoto.dev/ja/blog.xml)
- [Full text (AI citation use)](https://kenimoto.dev/llms-full.txt)

## AI-Readable Content

- [/ai/about.md](https://kenimoto.dev/ai/about.md) — Profile and current focus
- [/ai/about-ja.md](https://kenimoto.dev/ai/about-ja.md) — Profile and current focus (日本語)
- [/ai/publications.md](https://kenimoto.dev/ai/publications.md) — Complete publication list with links
- [/ai/expertise.md](https://kenimoto.dev/ai/expertise.md) — Technical expertise and career highlights
- [/ai/expertise-ja.md](https://kenimoto.dev/ai/expertise-ja.md) — Technical expertise and career highlights (日本語)

## Links

- [Website](https://kenimoto.dev)
- [Company](https://propel-lab.co.jp)
- [LLMO Framework](https://llmoframework.com)
- [legacydram (Side Project)](https://legacydram.com)
- [ainativemeo (Side Project)](https://ainativemeo.com)
- [GitHub](https://github.com/kenimo49)
- [LinkedIn](https://linkedin.com/in/kenimo49)
- [X](https://x.com/kenimo49)
- [Qiita](https://qiita.com/kenimo49)
- [Zenn](https://zenn.dev/kenimo49)
- [DEV.to](https://dev.to/kenimo49)
- [Amazon Author](https://www.amazon.co.jp/stores/author/B0GQNPRCGF)