Archive
What I've learned shipping things
41 articles across 3 topics. Real lessons, not theory.
Build Notes
Short lessons from real projects: code reviews, audits, AI mistakes, SEO gaps, and shipping decisions. Practical notes for founders and developers.
Open Build Notes- I Built 4 SaaS Products With AI. Total Revenue: $279. Creation got commoditized.
- Entity SEO Audit: Why AI Search Couldn't See My 5 Sites I ran 5 sites and 58 articles; to crawlers and AI models I didn't exist.
- AI Can Create Convincing Justification for Expensive Bad Decisions AI can produce correct evidence, polished reasoning and the wrong decision.
- How AI Tackles Tasks Differently Than Humans Experience lets you shrink a problem before you solve it.
- Does Claude Watermark Text? Yes. Do Nothing, Here's Why Claude is rolling out invisible text watermarks.
- Do Startup Directories Drive Traffic? 4 Products, 90 Days One product got zero referral sessions from 18 directories.
- Every Child of a Product Needs Its Own Moat Clean content isn't enough.
- Claude Design Brand Kit: GitHub Repo to Live Page in a Night I tested Claude Design the night it launched: a brand kit from a GitHub repo in 40 minutes, then a landing page live in production before morning.
- Automate Cloudflare Cache Purge + Warm After Deploy (Bash) Manual Cloudflare purges and cold first hits cost me 30 minutes a day.
- 5 Structural Patterns That Survive Every Code Review File-by-file review reads each file in isolation and calls it fine.
- AI Trusts Your Docs. That's the Problem. AI validates bad assumptions when you hand it bad ground truth.
- Audit Before You Spread the Mess Working code can still be wrong.
- Don't Assume AI Fixes Things Properly AI coding tools are good at text manipulation.
Claude Code
Working notes from daily Claude Code use: why it ignores CLAUDE.md, pre-commit hooks that block AI drift, skills vs scripts, and session habits.
Open Claude Code- AI Change Control: Keeping AI Code On-Architecture A tight CLAUDE.md and custom skills still let Claude Code ship 104 violations into my Laravel codebase.
- Three Weeks of Claude Code on a Raspberry Pi 5. The Login Held. My Temp Folder Didn't. The follow-up to my Raspberry Pi 5 Claude Code box.
- I Turned a Raspberry Pi 5 Into a Claude Code Box. Three Things Broke First. The cheap version of the Mac mini + Claude setup: a Raspberry Pi 5 running Claude Code headless on a subscription.
- Claude Code Best Practices: 6 Rules With a Receipt Six Claude Code habits, each traced to the failure that forced it: a 2292-to-16 marker check, a make-it-fail gate test, a landmine you size before you fix.
- Claude Code Commands: The Decision Rule From 42 in Use One markdown file, one description line, a name that becomes the command.
- Claude Code Memory: 4 Types, a 200-Line Cap Claude Code memory means two separate systems now.
- Claude Code Subagents: Delegating Is 3x Cheaper Which tier of Claude Code subagent to use for which job, backed by measured cost data: delegating is ~3x cheaper per turn, not more.
- AI Second Opinion: Why Your Own Agent Can't Give One I settled a decision with my project-manager agent.
- Claude Code Repeats the Same Mistakes: Memory Won't Fix It I built the memory file.
- I Didn't Do the Math on Max 20x. It Stopped Me Working. Every Pro versus Max guide runs the arithmetic.
- The Same Job Prices Out at $452 and $47 What it costs to run Claude over a real job.
- Three Agents, One Repo, Nobody Steps on Anybody Running several Claude Code agents against one repository without them colliding, and the four setup gaps that make the default arrangement quietly wrong.
- The Plan I Wrote in April and Never Opened Again A markdown build plan that went stale the week it was written, the four ways it failed once agents were doing the work, and the issue trees that replaced it.
- Claude Code Hooks: Make the Agent's Behavior Deterministic Skills guide Claude.
- Claude Code Skills vs Scripts: When to Use Each A skill guides the AI's judgment.
- Pre-Commit Checks for AI Code: A Zero-Dependency Hook Build a pre-commit hook that stops AI code drift before it commits.
- Laravel Pre-Commit Hook Audit: 104 CLAUDE.md Violations A grep-based pre-commit hook found 104 violations in a Laravel codebase whose CLAUDE.md already banned them.
- Why Claude Code Ignores Your CLAUDE.md (and the Fix) Claude Code treats CLAUDE.md rules as suggestions, not walls, so it skips them.
- Fable 5 as the Driver: When to Let One AI Instruct Another I used Claude Fable 5 to drive Opus across five live projects: a brilliant strategist and a dangerous judge.
- What 17 Retrospectives Taught Me About Coding With AI I ran a retro after every build for three months.
- Stop Claude Code From Deferring to 'Next Session': One Rule Claude Code keeps deferring work to a 'next session' that never comes.
- From 733 False Positives to 9 Safe Deletes with Claude Code 733 dead-code candidates, 90% false positives.
Research
Measured findings from running AI coding agents in production. Every post here has a denominator, a published dataset, and its own errata.
Open Research- I Pulled Apart a Brand With 80 Facebook Pages. The Lesson Wasn't About Ads. A brand that runs about 80 Facebook pages for one product, and barely mentions the product.
- Agent Notes About Open Bugs Go False. Fix Stories Don't. I checked 377 notes my coding agents wrote about files against today's code.
- I Counted Every Time Claude Code Auto Mode Said No I parsed six weeks of my own Claude Code transcripts for auto mode classifier denials: 2.3 per 1,000 tool calls, 13 of 71 rules, and a script to run on yours.
- I Build Analytics Tools. Mine Drew Me a Backwards Chart. The default scroll report said deeper scroll converts better.
- My Bounce Rate Was 91%. One Robot Did That. A high bounce rate on a small site is often machines, not people.
- Your Coding Agent Has Memory. Does It Actually Arrive? I instrumented my own agent memory system and measured a 24.6% delivery rate, then found two bugs in the instrument itself.