Each text is a real problem that showed up mid-work — and the tool that now keeps it from coming back.
How to gate an AI coding agent with an independent critic: a 0-100 score backed by real commands, an independent reviewer and a fail-closed security gate. Plus the four ways a green loop lies.
Read it →See the skillAn OAuth-protected MCP server reports connected, lists its tools and returns 401 on everything account-scoped. Four curl commands to diagnose it and the fix that makes the authorization flow actually start.
Read it →See the skillA screenshot test fails for two opposite reasons — the page changed, or the capture is not reproducible. The four verdicts (PASS/FAIL/STALE/ERROR), the budgets that survive contact with reality and the determinism checklist.
Read it →See the skillAn MCP server that answers on curl but never comes up in the client: a short connect timeout, a missing API key header, a config read at import, and a session that died mid-call. How to measure each one.
Read it →See the skillA watchdog that reports all green: a process counter that counts itself, a fixed resource gate, a lost delivery looping. How to prove a guard works — with a negative test, not with hope.
Read it →See the skillA visual test that passes locally and fails in CI: date, time, counter, font and animation. How to make the capture deterministic before blaming the diff.
Read it →See the skillA block page, a cookie wall and a login redirect all return 200. A count that does not add up, a hostname pointing at another app, and the difference between unverifiable and verified.
Read it →See the skillScraping supplier pricing to build a margin: confirm which side of the table you are reading, add the real shipping cost, and pick the SKU by ticket size, not by percentage.
Read it →See the skillA checkpoint that fails silently: wrong interpreter, sanitised path, locked database. How to prove it wrote — by reading it back — instead of trusting the exit code.
Read it →See the skillMeasured distribution: Reddit removes a new account's post, X blocks on a character limit that looks like a ban, dev.to publishes but canonical needs an API key. What to do on each channel.
Read it →See the skillAn agent that confirms what already existed and calls it a delivery. How to separate “I checked” from “I produced”: a new artifact, execution evidence, and the check that fails when nothing was created.
Read it →See the skillA scheduled job that reports success without delivering: platform not configured in the running process, manual vs scheduler execution, and the delivery id that separates green from false green.
Read it →See the skillLint and tests that pass locally and fail in CI: an unpinned dependency, a different environment, an unversioned file, and a cache that lies.
Read it →See the skillThe agent that writes the code and scores itself always scores high. The critic has to be another process: a real command, a 0-100 score, and the reason when it fails.
Read it →See the skill