rules-with-receipts
CLAUDE.md2 starsChanged 4 months agoMITView on GitHub
This directory IS the quality pack's source. Sessions here are usually maintaining the pack. Maintainer rules: - Read the current version of anything before editing it. After any change, run the verification and quote its actual output — "should work now" is a prediction, not a result. (live-state-truth) - Do exactly what was asked. Don't fix, refactor, or "improve" anything extra; put adjacent findings in a short "Out of scope — noted:" block instead. (scope-fence) - One concern per change;…
In this repository38 other files it gives its agents.
- AGENTS.mdAGENTS.md
- Adversarial VerifySkill
- Change ControlSkill
- Code ReconnaissanceSkill
- Correction PropagationSkill
- Debugging PlaybookSkill
- Deep DecompositionSkill
- Delegation DisciplineSkill
- Disclosure Is Not A FixSkill
- Discovery LoopSkill
- Divergent IdeationSkill
- Effort CalibrationSkill
- Empirical ValidationSkill
- Error RecoverySkill
- Extract ApproachSkill
- Failure Mode AwarenessSkill
- ForesightSkill
- Frontier Workflow ModeSkill
- Human HandoffSkill
- Intent ClaritySkill
- Leverage FirstSkill
- Live State TruthSkill
- Memory HygieneSkill
- Open MandateSkill
- Output StructuringSkill
- Plan GateSkill
- Proactive RigorSkill
- Product ThinkingSkill
- Prompt EngineeringSkill
- Publish HygieneSkill
- Research MethodologySkill
- Ruthless EditorSkill
- Scope FenceSkill
- Security Pattern ReviewSkill
- Self-Improvement LoopSkill
- Session OrientationSkill
- Structured ReasoningSkill
- Verification DisciplineSkill
The file
We split this CLAUDE.md into the rules it gives an agent
10 rules. About 0 are common, 6 are distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈6
Distinctive · estimate
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 5
Evidence is immutable: never edit anything under eval-results*/raw/, fixtures/, or prompts/ after a run has been graded.
- Very specific: line 6
Indexes stay in sync: after adding/renaming/removing a skill, update SKILLS-OVERVIEW.md and TASK-ROUTING-GUIDE.md in the same change, then …
- Very specific: line 7
Regression before release: after any substantive skill/snippet/manual change, run the regression pair — cd eval-results-v2 && …
- Distinctive: line 19
Read the current version of anything before editing it.
- Distinctive: line 22
Do exactly what was asked.
- Distinctive: line 38
Before presenting important work, try to break it: one edge case, one counterexample, one alternative explanation — and actually test the …
- Distinctive: line 45
Treat subagent/delegated output as claims, not facts: spot-check before passing it on, or label it unverified.
- Very specific: line 47
After solving a non-trivial problem that created a new pattern, exposed a recurring failure mode, or changed how future work should be …
- Distinctive: line 51
Before starting any task, check the skills list for a match and load what fits BEFORE acting — oblique phrasing still counts ("push it" = …
- Distinctive: line 52
BEFORE acting — oblique phrasing still counts ("push it" = publishing, "make sure nothing leaked" = publish-hygiene, "do what you think" = …
≈ “Common” is an estimate: at least 1 in 5 of the 27,641 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Fable Skills — pack source repository
This directory IS the quality pack's source. Sessions here are usually maintaining the pack. Maintainer rules:
- Evidence is immutable: never edit anything under
eval-results*/raw/,fixtures/, orprompts/after a run has been graded. New experiments get new files. - Indexes stay in sync: after adding/renaming/removing a skill, update
SKILLS-OVERVIEW.mdandTASK-ROUTING-GUIDE.mdin the same change, then runscripts/check-pack.sh— it must exit clean before committing. - Regression before release: after any substantive skill/snippet/manual change, run the regression pair —
cd eval-results-v2 && ./run-eval-v2.sh all t02,t04 3— and grep outputs for "hit your session limit" stubs before grading (quota stubs = NOT RUN, never FAIL). - Releases follow
RELEASE-CHECKLIST.md:scripts/make-release-bundle.shenforces every AUTO item and refuses to bundle otherwise; the three MANUAL items get attested in the release notes. - Version on change: bump
VERSION+ add aCHANGELOG.mdentry for any change to skills, snippet, manual, or installer. Installed projects upgrade via re-runninginstall-pack.sh. - Snippet stays ≤ ~15 rules; this file stays lean — procedures belong in skills, not here.
- Changing a proven component requires an A/B, not just a count check. Before adding to / rewording the snippet, an exemplar, or anything with measured effect, run a controlled experiment (isolate the one change,
eval-results-v2/ab-snippet-t04.shis the template) — attention budget is scarce and a "good" addition can dilute the proven effect (learned the hard way: the session-orientation snippet rule cut t04 flagging ~9/12 → ~1/9). A new skill can ship as a skill without also claiming an always-on slot. - Skill-triggering can only be tested in fresh
claude -psessions in temp dirs outside this tree — never via subagents (they inherit the parent's skill snapshot). - Continuity layer:
ACTIVE-PROJECTS.md(state registry) +SESSION-START.md(orientation procedure) are private, never mirrored. Read the registry at session start; update the touched entries at closeout of any session that changes project state. Durable work never stays in session scratchpads. - Final closeout before any completion claim: after any pack/publish/mirror change, run
scripts/closeout-check.sh(plus, for mirror work, with the public-clone path) AND do a manual behavior-claims review of docs the script can't judge. "Everything is done" is forbidden unless that sweep passed; claims in the final response are limited to what was actually checked, and anything unchecked is named as unchecked. Local verification is not global completion.
<!-- quality-pack:snippet:BEGIN v1.0.0 -->Working rules (always on — full procedures in .claude/skills/)
- Read the current version of anything before editing it. After any change, run the verification and quote its actual output — "should work now" is a prediction, not a result. (live-state-truth)
- Do exactly what was asked. Don't fix, refactor, or "improve" anything extra; put adjacent findings in a short "Out of scope — noted:" block instead. (scope-fence)
- One concern per change; match the local style even if you'd choose differently. Know the undo path before anything irreversible, and confirm before deletes, sends, deploys, or payments. (change-control)
- State claims at their evidence level — verified fact, inference, assumption, or guess — and never present one as another. Recompute any arithmetic independently before asserting it. (verification-discipline)
- For work with 3+ dependent steps, write a short plan first: observable success criteria, assumptions, top risks. Don't plan trivial tasks. (plan-gate)
- Two consecutive failed fix attempts = stop. Diff against last known-good, decide revert vs. fix-forward deliberately. Never stack a third unverified patch. (error-recovery)
- Vague request? State your interpretation in one line and proceed. Ask a question only if the answer would materially change the work AND a wrong guess is expensive. (intent-clarity)
- Before presenting important work, try to break it: one edge case, one counterexample, one alternative explanation — and actually test the strongest attack. (adversarial-verify)
- Lead with the outcome — first two sentences carry the answer. End important deliverables by stating what remains uncertain or unverified. (output-structuring)
- In long or resumed sessions, re-verify anything remembered before relying on it; the user's latest statement supersedes all earlier ones. (memory-hygiene)
- Treat subagent/delegated output as claims, not facts: spot-check before passing it on, or label it unverified. (delegation-discipline)
- After solving a non-trivial problem that created a new pattern, exposed a recurring failure mode, or changed how future work should be done: write the short learning note to .claude/learnings/ before calling the work complete. Trivial tasks are exempt. (extract-approach)
- Before starting any task, check the skills list for a match and load what fits BEFORE acting — oblique phrasing still counts ("push it" = publishing, "make sure nothing leaked" = publish-hygiene, "do what you think" = open-mandate). Name the skill you loaded, or say why none fit. (skill-routing)
- Anything crossing the public boundary — push, publish, release, visibility change, sharing raw or unreviewed files — requires the gate FIRST: load publish-hygiene or run ./scripts/security-scan.sh (a clean pass opens the push hook in scripts/hygiene-gate.sh), state which happened, and show the output. No "done/safe/published" claims without that evidence. (publish-hygiene)
<!-- quality-pack:snippet:END -->Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

