cd /news/ai-tools/a-rulebook-rots-quietly-what-checks-… · home › topics › ai-tools › article
[ARTICLE · art-170313] src=dev.to ↗ pub= topic=ai-tools verified=true sentiment=↑ positive

A rulebook rots quietly. What checks it every Monday?

Omer Bin Saleem, a frontend chapter lead, built a scheduled read-only lint that runs weekly against a team's Claude Code rulebook to catch contradictions and stale entries. The job combines a plain script for broken IDs, commands and links with four separate claude -p passes (CONFLICT, STALE, DRIFT, DUPLICATE) restricted to Read, Grep and Glob, run via launchd every Monday at 11:00 from remote main. The first report surfaced 41 findings, three of which were confirmed by hand as real, including a gate script that blocked a push the handoff guide said goes through.

by read2 min views4 publishedNov 16, 2026

Originally published at omerbinsaleem.com by Omer Bin Saleem (Omer Saleem). Rules for Claude Code pile up until two of them contradict each other and nobody notices. A scheduled read-only lint found the contradictions the team had stopped seeing.

#

How to do it

  1. Start without a model. A plain script can find rule IDs, commands, scripts and links that point at nothing. Those are the cheapest findings and they need no Claude at all.
  2. Add four read-only passes with claude -p, one question each: CONFLICT, STALE, DRIFT and DUPLICATE. Give it only Read, Grep and Glob, no project settings loaded, no session saved. One pass with one question is easier to trust than one pass that asks everything.
  3. Put it on a schedule, not in a hook. Ours is a launchd job on Monday at 11:00. It runs the lint script straight from the remote main branch, so the local checkout is never pulled and never changes under it.
  4. Write the report to a dated file and keep a log. The first report held 41 findings. Three were checked by hand and were real: a gate script blocked a push that the handoff guide said goes through, and a guide accepted a ticket prefix that a check script refused. Each had sat there while people read the files every day.
  5. Keep the fixes human. Nothing in the job edits a rule. Every finding goes through the same review as any other rule change, because only one person on the team approves rules.
  6. Say no to the full wiki. We looked at the LLM Wiki pattern and took only the lint job. An earlier auto-capture queue had collected 1669 events and nobody ever read one of them. A weekly report someone opens beats a stream nobody consumes.

#

Prompt to copy

I am Omer Bin Saleem, a frontend chapter lead who builds with AI every day. One practical AI hack a day at omerbinsaleem.com. LinkedIn: omer-bin-saleem · Instagram: @iomerbinsaleem.

Written with AI assistance, from my own work notes.

── more in #ai-tools 4 stories · sorted by recency
── more on @omer bin saleem 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/a-rulebook-rots-quie…] indexed:0 read:2min 2026-11-16 · —