Notes from the workbench.
Thinking out loud about PHP, AI tooling, ABA software, indie hacking, and whatever else I'm shipping this month.
A Floor, Not a Ceiling: Why My Agent Skills Use Everything Claude Code Offers
Most of my agent skills started as Matt Pocock's. His are written to run the same in every harness. Mine still run anywhere, but where Claude Code has a better tool for the job, they use it. Here's what that changes when you're the one sitting at the keyboard.
Six Rounds of Bot Review on One PR
An agent built a whole feature in one session: interview, spec, design, code, tests, a live click-through. It passed its own review. Then the review bots found 32 real bugs, one round at a time. Asking their questions before the PR is the fix.
When 'Always' Means Always: The Implied Clause in Your Agent Rules
I told my agent to ALWAYS add --tia to Pest. It did — by running the whole suite after every single step. A human hears an implied 'when you run Pest.' The agent doesn't. Here's the lesson.
Multiple Claude Code Accounts, One Shared Setup
Claude Code can isolate each account in its own config directory, but that usually leaves every profile with stale settings and missing skills. Symlinks give each account the same personal setup without sharing its identity.
I Went on Vacation and My Agents Paid Down My Tech Debt
I took a week off and let Superconductor loose on a 500k-line Laravel codebase. It opened roughly two dozen PRs a day: N+1 fixes, unbounded job processing, missing HTTP timeouts, tenant-scoping bugs. I merged 98% of them. This is the setup that made that safe.
Turning a Creator's X Tips Into Agent Rules Claude Actually Reads
I follow Povilas Korop for Laravel tips, but they scroll by on X and vanish. So I scraped ~696 of them, OCR'd the code-in-image samples with Claude, and distilled the durable ones into versioned agent-rules repos my coding agents load automatically. Here's the whole pipeline.
My Global CLAUDE.md: The Rules I've Built Up Over Time
Every time an AI agent burned me, I wrote a rule. Here's my entire ~/.claude/CLAUDE.md — the personal rules that travel with me across every project — and the scar behind each one.
A Linear workspace per project in Claude Code (no more reconnecting)
If you work across two projects with two different Linear workspaces, Claude Code's Linear MCP makes you disconnect and reconnect every time you switch. Here's how to bind each repo to its own Linear account so it just works.
Stop stuffing every AI rule into one root file: apply-agent-rules
Most projects pile every AI agent rule into a single root CLAUDE.md and the file balloons past a thousand lines. Subdirectory rule files are a much better idea — and apply-agent-rules makes them practical.
One LLM Isn't Enough: The Data
Two months of multi-model code review across 217 commits and 1,514 findings. Consensus is rarer than I thought, and 87% of real bugs were caught by exactly one model.