CLAUDE.md patterns that actually bind

September 12, 2026 · View on GitHub

Install: npx skills add soul-sol/claude-md-patterns

Your CLAUDE.md probably says "write clean code, be careful with destructive operations, ask if unsure." The agent reads it, agrees with it, and then does whatever the last 200 lines of context suggested. The file isn't wrong — it just isn't a rule. Nothing in it can be checked, so nothing in it binds.

A rule that binds has four parts:

  1. Trigger — the exact condition where it applies ("any browser launch", "before every commit", "when a worker exits")
  2. Check — a command, string, or state that is either true or false. If a reviewer can't grep for it, the agent can't enforce it on itself
  3. Stop condition — what to do when the check fails, named explicitly (ABORT: …), not implied
  4. Evidence — what must be left behind so the rule's application is visible afterward

These came out of running Claude Code and Codex workers in parallel every day. The incidents behind them — what the agent claimed, what actually happened, and the gate that catches it next time — are at status.lifestep.io/incidents. The full set is The CLAUDE.md Pattern Library ($9), or the Complete Agent Ops Kit ($29) if you also want the task briefs, the 25 review prompts and the deploy gate that go around these rules; what is here stays free and MIT either way.

Audit your own file first

Before adopting any pattern here, see which of your existing rules an agent cannot act on:

claudemd.lifestep.io — paste a CLAUDE.md or AGENTS.md and it flags truncation risk, headings with no rule under them, instructions with no verification command, prohibitions with no stated alternative, and rules that read like controls but enforce nothing.

It is a static checker, not a model: no upload, no API call, no LLM. The page runs the checks in your browser and cites the line each finding came from. Checks it cannot evaluate report not evaluated rather than pass — absence of a shape to inspect is not evidence that the file is fine.

5 free patterns

Each file gives the misbehavior it fixes, the exact text to paste, why it binds, variations (monorepo/solo/team), and the anti-pattern — the plausible version that fails.

#PatternStops
05Autonomy clauseworkers that stall waiting for approval nobody will give
10Done is evidence"it's finished" reports that were never verified
16Browser profile quarantineautomation touching your real browser profile (this one came from losing ~2,000 bookmarks)
19Look before you destroyoverwrites and deletes on the wrong target
21Reassign, don't retryburning three attempts on a worker that will fail again

Adopt these five first. Add more only when you hit the specific failure they address — a rule adopted after the incident gets followed; a rule adopted preemptively becomes wallpaper.

Test that a rule binds

Open a fresh session (new process, no history) and probe it:

# 1. Can it restate the rule unprompted?
claude -p "Without looking anything up: what must you do before launching a browser in this project, and what is the abort condition?"

# 2. Give it a task that would violate the rule if the rule were absent.
claude -p "Take a screenshot of my inbox for the report."

Vague answer to (1) → the rule is too long or too far down the file. (2) proceeds without the guard → the rule is not binding. Tighten trigger and stop condition, probe again. A rule you have never probed is a rule you are guessing about.

The full library (30 patterns)

Same format as the five above — each pattern names the failure it stops:

PatternStops
Pipefail standarda failing test reported as a pass, because $? came from the pipe's last command
Deploy success markerreading a rollback's "health checks passed" as a successful deploy
Warning is not failurepulling a working worker out of rotation over an informational log line
Stall doctrinecalling a finished job stalled because it didn't print the marker you expected
Polling-loop trapwatch loops that die on a timeout and take the run's only observer with them
Per-worker signalsa finished job marked stalled because it lacked another tool's completion marker
Context budgeta correct rule that loses to the last 200 lines of context
Precise process killsa cleanup that kills more than the thing you meant

Plus 17 more beyond the free five — global/project split, subdirectory overrides, worker exception clause, brief contract, small-diff rule, file ownership map, worktree mandate, log-tail check, review gate before commit, identity actions human-only, secret existence checks, memory file discipline, session parking, progress report format, model tier policy, bulk-read delegation, latest-model pin — and the five above in the same format, with a 43-page PDF adoption guide, the probe test, and a Korean guide.

→ $9 on Gumroad

Rules decide what an agent may do. They don't tell you what to hand it, or how to judge the diff that comes back. The Complete Agent Ops Kit ($29) covers the whole loop — this library in full, plus task-brief and change-scope templates, 25 adversarial review prompts to run against a worker's diff before it lands, the deploy/rollback gate, and Solo, Like a Team in English and Korean. The same files bought as separate products: $49.

License

MIT for the 5 patterns in this repository.

The paid guide collection is available at lifestep1.gumroad.com.