Blog
Articles on AI development, LLMO, context engineering, and harness engineering.
2026 101 posts
- Codex CLI vs Claude Code: 7 Real Tasks, Same Repo, 31 Days Later 17 min
- GitHub Copilot Agent Mode vs Claude Code: 8 Tasks, 31 Days, Real Bills 17 min
- Claude Code vs Aider on Same 10 Tasks — One Finished 8, the Other 5 14 min
- Claude Code vs Qwen3-35B on RTX 4070: 34.6 tok/s Break-even 17 min
- Karpathy's nanochat on a MacBook: End-to-End runcpu.sh Log from Tokenizer to Chat (2026) 13 min
- Claude Code Has Two Authentication Layers. setup-token Is Not the Same as login. 12 min
- The Shai-Hulud npm attack: why signature verification, npm audit, and --ignore-scripts all failed 8 min
- I stopped making Claude Code draw my diagrams and started drawing with it 11 min
- Claude Code: The Operator's Field Guide 16 min
- llms.txt vs Claude Skills manifest: I tested 5 AI engines to see which file they actually read 18 min
- When Claude Code's Auto Mode Blocks Only Bash: Investigating the Safety Classifier Outage, Plus a Fail-Open-on-Outage Hook Design 16 min
- Chat Models Are Born in the Loss Mask — Reading nanochat's SFT 11 min
- Five Days of Tracking 346 Aftershocks From 90 km Away — and a Rematch With the 2016 Kumamoto Earthquake 7 min
- Karpathy nanochat: How GPT-2-Class Training Fell From $43,000 to $48 in 7 Years (2026 H100 Spot Math) 15 min
- Why Earthquake Prediction Is Impossible but Aftershock Forecasting Works: Measured on 130 Noto Aftershocks 23 min
- llmoframework Audit on 30 Dev Blogs: 4 of Top 5 Fail Same 3 Checks 20 min
- Cursor Composer vs Claude Code: 400k-Token Repo Benchmarked 11 min
- Karpathy Nanochat: $43K GPT-2 Training in 8,000 Lines (2026) 12 min
- PinchTab Only Shoots One Viewport: Teaching a 9.4k-Star Browser Bridge to Capture Full Pages 10 min
- pixelshot Read One Tile of an 18,609px Wikipedia Page: the Lazy-Load Trap Under Visual RAG 14 min
- The Machine Accent Travels: AI Text Is Rhythmically Monotone in 70/70 Cells Across 3 Languages 7 min
- OpenAI Codex AGENTS.md: 1M Lines Shipped, 3 Harness Lessons 11 min
- Article Schema Alone Didn't Make AI Recognize Me as the Author. The Entity Wiring That Did (in 4 JSON-LD Fields). 17 min
- I Forked OpenCut classic to Put an MCP Server on It — Four Traps I Hit Before Claude Could Drive the Video Editor 13 min
- Pre-flight your MCP: four layers to grade a server before you publish it 24 min
- The Skill Eval Repo I Didn't Build: 107 SKILL.md Files, 6 Checks, 21 False Positives 15 min
- Ship a product, get a support button for free: an edge-injected overlay on Cloudflare Workers 12 min
- FUNDING.yml alone won't show a Sponsor button: notes from auditing 42 repos 8 min
- historymap: one YAML file becomes a corporate-style product-history timeline 9 min
- I Trusted Claude Code and Shipped 40% Slower: 3 Places the Speed Actually Died 18 min
- The 7-Step Test That Told Me When to Switch From RAG to GraphRAG 18 min
- I Ran Claude Code, Cursor, and Codex Side by Side for 31 Days. Here Is the Real Monthly Bill. 20 min
- Claude Code vs Cursor: 6 Tasks, Measured Latency, and Which One I Uninstalled 17 min
- Parallel Agents Went Negative at 340k Tokens: The Real Breakeven for Claude+Cursor+Codex 16 min
- RTX 4070 + One llama.cpp Flag = 2.8x Tokens/Sec (Ollama Default Is the Loser) 16 min
- Your AI Code Review Burns 80% of the Context Window on Files It Never Needed 19 min
- LLMO: The Field Guide to Getting Cited by AI Search 20 min
- I Made Claude Code Review Only the Blast Radius — Token Bill Dropped 8-49x 16 min
- Anthropic Just Rewrote Their frontend-design Skill — and Named 3 AI Design Clichés (With Hex Codes) 19 min
- I Ran Claude Code, Cursor, and Codex in Parallel for a Day. The Real Cost Was 412 Decisions. 19 min
- Measuring AI Citation Half-Life: A 90-Day Methodology With 3 Real Decay Curves 21 min
- Multi-Agent Decision Fatigue: I Counted 412 Micro-Choices a Day. The Harness Cut It to 38. 17 min
- I Added 3 Numbers to One Paragraph. Perplexity Started Citing It in 11 Days. 22 min
- AI Mode Just Hit 1 Billion Users, and Opened a Local-Business LLMO Market Most Engineers Are Ignoring 13 min
- I Wired My Pages Into Topic Hubs, Not a Flat List: AI Citations Consolidated Onto 4 of Them 14 min
- AI Search Is Under 1% of My Traffic and 12% of My Signups. That's the LLMO Case I Actually Use. 15 min
- GEO's +115% From Statistics Is Domain-Dependent: It Worked for My Tech Posts and Did Nothing for My How-To Pages 14 min
- AI Reads Your Chunks, Not Your Page: I Promoted 9 Sections from H3 to H2 and Watched Which Ones Got Quoted 18 min
- I Wired Claude Code to Real Hardware Over USB Serial. The MCP Tool Was the Easy Part. 18 min
- I Checked What GPTBot Actually Sees on My JS-Rendered Pages. It Was an Empty `<div>`. 13 min
- AI Wrote 100 Passing Tests. Mutation Testing Says They Caught 58% of Real Bugs. 12 min
- AI Finds Your Page Three Ways. I Published the Same Fact in All Three and Timed Which Reached AI First. 14 min
- I Priced AI Agents Three Ways: API, Subscription, and Local. Here's Where the Break-Even Actually Sits. 14 min
- My Blog Speaks 4 Languages. AI Search Cited the Wrong One to Most of My Readers. 12 min
- I Rewrote 12 Pages to Answer the Question in the First Sentence. AI Started Quoting 7 of Them. 13 min
- I Rank #1 on Google. On Brave I'm Page 5. My Own AI Agents Can't Find Me. 18 min
- Perplexity Citations Exploded After I Changed 3 Things. Only 1 Was Schema. 20 min
- I Gave Every Page on My Site a .md Twin. The AI Fetchers Stopped Guessing 13 min
- My Best Page Went Stale in a Month: Why AI Search Rewards Freshness, Not Just Schema 12 min
- AI Search Splits Your One Question Into Six. My Pages Answered None of Them. 14 min
- I Stopped Adding Context to My Agent and Pruned Tool Outputs Instead — My 3-Hour Task Stopped Forgetting Its Own Plan 13 min
- AI Citations Have a Half-Life. I Tracked Mine for 9 Weeks and Watched Them Decay. 16 min
- I Mapped My Codebase as a Graph. The File That Broke Was Two Hops Away. 15 min
- Link-less Brand Mentions Beat Backlinks for AI Visibility — I Read the Ahrefs 75,000-Brand Study So You Don't Have To 15 min
- Your Page Rank Is Invisible to AI — Only Your Passages Get Cited 15 min
- I Crosspost to 4 Platforms with rel=canonical Pointing Home. AI Search Still Picks the Copy. 16 min
- Claude Code Skills Cost Tokens Even When They Don't Fire. I Measured 5 Skills Across 7 Hours. The Bill Was 18%. 17 min
- I Cron-Scheduled 7 AI Agents. 2 Silently Failed for 18 Days. Tracing Wouldn't Have Caught It. 19 min
- I Ran 3 Claude Code Sessions in Parallel for 8 Hours. They Overwrote Each Other's Context Twice. 18 min
- I Asked 5 AI Search Engines to Cite My Own Blog. Only 3 of 31 Articles Showed Up. 16 min
- I Added 11 JSON-LD Schemas. Three Months Later, Only 3 Showed Up in AI Citations. 18 min
- I Refactored 100 Functions With Claude. 7 Got Slower in Production. 17 min
- I Told Claude Code to Do TDD. It Wrote the Test AFTER the Code 6 Out of 10 Times. 20 min
- I Added a 4th Agent That Audits My Other Agents. It Caught My Strategist Procrastinating for 3 Weeks. 23 min
- I Translated My Blog Into 4 Languages. Portuguese Got Nearly 4× the Traffic of English. 14 min
- TRM's 8,337% LLMO Playbook on Indie Sites: Only 1 of 4 Pillars Worked 19 min
- Claude Said 'You're Absolutely Right!' 47 Times Last Week. I Was Only Right 11 Times. Claude Was Wrong 36. 12 min
- I Plugged the Same Site Into 7 AI-Citation Trackers. They Reported 7 Different Numbers. 15 min
- The 5 AI Crawlers That Hit My Sites Most in 30 Days — What Their Logs Told Me About LLMO 15 min
- I Plugged Claude into a Chaos Engineering MCP Server. It Killed Staging 4 Times Before Finding a Bug We'd Missed for 6 Months. 31 min
- I Caught Claude Hiding My Bug 3 Times in a Row. Then I Turned 10 Debugging Habits Into Prompts. 18 min
- I Gave My Strategist Agent WebSearch. 5 Topics Took 20 Minutes. Splitting It Into 3 Made It 3. 15 min
- I Benchmarked 5 Voice AI Stacks. Only 2 Stayed Under 300ms. 16 min
- 3 Claude Code Sub-agents Reviewed One PR — They Disagreed on 40% 18 min
- I Audited 30 llms.txt Files in the Wild. 5 Anti-Patterns Are Already Forming. 19 min
- OpenClaw Hit 250K Stars Faster Than React. I Spent a Day Switching From Claude Code 18 min
- I Refused to Write Specs Until Claude Code Generated Wrong Code Three Times 15 min
- I Let My Claude Code Agent Run for 24 Hours. The $400 Bill Was the Least Scary Part. 19 min
- The og:type Bug Three of My Astro Sites Quietly Shipped 11 min
- I Stacked 4 More Context Layers on Top of RAG. The Improvement Was 12%. 14 min
- Natural-Language Agent Harnesses (arXiv 2603.25723): 4 Concrete Patterns from Reading the Paper Twice 22 min
- Claude Code Skills: The Reusable Workflow That Replaced My Commands 17 min
- Your New Domain's First Week of GA4 Is a Lie: 4 Days of Raw Data from kaoriq.com's Launch 11 min
- ChatGPT Codex vs Claude Code: 6 Agent Tasks, $297 API Bill, Which One I Kept in 2026 22 min
- One Question, Five AI Search Engines, Five Different Answers 19 min
- Is AI Actually Citing Your Site? How to Measure What Google Rankings Can't 16 min
- Princeton Tested 9 Ways to Get Cited by AI. Only 3 Worked. 18 min
- 9 Bugs in My AI Pipeline: None Were the AI's Fault 9 min
- The Cheap Model That Won: Why Context Beats Parameters 12 min
- llms.txt: The File That Decides Whether AI Can Find Your Site 13 min
- Building an Autonomous Content Pipeline with Claude Code 2 min