l-gevity-skills
CLAUDE.md88 starsChanged 3 days agoMITView on GitHub
How an agent thinks about tasks in any codebase using this skill library. Skills in .claude/skills/ define what good code looks like; this file defines the attitude. Any skill is invocable by name — /<skill> in Claude Code, $<skill> in Codex — or by a request matching its description. Bias: caution over speed on non-trivial work; don't bureaucratize one-liners. M returns BUILD / KEEP / SIMPLIFY or stop. Only READY, or PARTLY-READY as a bounded reversible increment, enters A; NOT-GROUNDED,…
In this repository52 other files it gives its agents.
- alchemy-orchestrateSkill
- alchemySkill
- architecture-as-code-javascriptSkill
- architecture-as-code-pythonSkill
- architecture-as-codeSkill
- architecture-guidelinesSkill
- bring-downSkill
- ci-cd-reliability-architectureSkill
- continuous-improvementSkill
- defect-shift-leftSkill
- dependency-lifecycleSkill
- evolutionary-database-designSkill
- functionality-complexity-tradeoffSkill
- gate-lifecycleSkill
- implementation-readinessSkill
- infrastructure-fitSkill
- morphogenetic-architectureSkill
- observability-designSkill
- push-outSkill
- requirements-groundingSkill
- requirements-topologySkill
- requirements-traceabilitySkill
- structural-simplificationSkill
- system-optimizationSkill
- test-strategySkill
- zero-copy-requirementsSkill
- alchemy-orchestrateSkill
- alchemySkill
- architecture-as-code-javascriptSkill
- architecture-as-code-pythonSkill
- architecture-as-codeSkill
- architecture-guidelinesSkill
- bring-downSkill
- ci-cd-reliability-architectureSkill
- continuous-improvementSkill
- defect-shift-leftSkill
- dependency-lifecycleSkill
- evolutionary-database-designSkill
- functionality-complexity-tradeoffSkill
- gate-lifecycleSkill
- implementation-readinessSkill
- infrastructure-fitSkill
- morphogenetic-architectureSkill
- observability-designSkill
- push-outSkill
- requirements-groundingSkill
- requirements-topologySkill
- requirements-traceabilitySkill
- structural-simplificationSkill
- system-optimizationSkill
- test-strategySkill
- zero-copy-requirementsSkill
The file
We split this CLAUDE.md into the rules it gives an agent
4 rules. About 0 are common, 2 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 13
Read a module's exports and 2–3 nearest callers before extending it.
- Very specific: line 16
Verify the problem exists in this stack before step 1 of any prescribed fix.
- Very specific: line 78
Before writing prose, try to encode it as a lint check, type, test, or build-time gate — manual rules drift, encoded ones don't.
- Distinctive: line 88
If a sentence doesn't change what the reader does next, drop it.
≈ “Common” is an estimate: at least 1 in 5 of the 28,029 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
CLAUDE.md — Strategic Directives
How an agent thinks about tasks in any codebase using this skill library.
Skills in .claude/skills/ define what good code looks like; this file
defines the attitude. Any skill is invocable by name — /<skill> in Claude
Code, $<skill> in Codex — or by a request matching its description. Bias: caution over speed on non-trivial work; don't
bureaucratize one-liners.
1. Think before coding
State assumptions. Multiple readings → name them and stop. Confused → say what's unclear.
2. Read before you write
Read a module's exports and 2–3 nearest callers before extending it. Can't explain its shape → ask.
3. Necessity before execution
Verify the problem exists in this stack before step 1 of any prescribed fix. Authors prescribe;
you verify the prescription matches a real problem. → functionality-complexity-tradeoff §1
4. Match conventions; surface conflicts
Conformance beats local taste. Contradicting patterns → pick one (more recent or more tested), flag the loser, or escalate if migration cost is real. Never blend a hybrid.
5. Surgical changes
Touch only what the task requires. No adjacent improvements, no refactor of what isn't broken, no helpers for one-shot work.
6. Walk the adaptive pipeline in order
/alchemy, $alchemy, or a natural request such as "do some alchemy" runs the alchemy
skill, which owns routing, gates, handshakes, companion order, and the decision trail; this
section keeps only what every session holds. It dispatches before loading any gate: SKIP
routine local work, DIRECT one clear gate, ADAPTIVE structural work, and FULL only for
explicit full-traversal language. Focused aliases stay focused and report missing
prerequisites; a core skip never suppresses an independently matching companion skill.
Resume from the latest decision artifact whose decisions name what they supersede:
Requirements Grounding, when evidence or meaning is absent or stale
→ M — Minimum
→ Requirements Topology, when relationships are non-trivial
→ Implementation Readiness
→ A — Architecture → L → C → E → H → YM returns BUILD / KEEP / SIMPLIFY or stop. Only READY, or PARTLY-READY as a bounded
reversible increment, enters A; NOT-GROUNDED, BLOCKED, and NOT-READY stop or return to
the failed stage. Audits start at the read-only C₀ structural baseline. Y optimizes a
stable, measured baseline, so it runs in the iteration after the increment ships.
A subsystem is where change lands; an aspect, one obligation and one mechanism across declared subsystems, is which dimension is touched; an increment is what changes; an iteration admits, realizes, and measures an increment, and iteration 2 is the next cycle on the same subject, starting from that measured baseline.
- Enforcement files for a new subsystem ship in the same PR as its code; spike code is the only exception and never crosses the merge boundary without rules.
- Expand and contract never ship in one deployable, and the contract step is gated on evidence, not a date.
- Completion proves a working capability, not downstream impact.
READYnever means implemented; a code anchor never means verified. zero-copy-requirementsplaces each fact: code and tests own behavior, history owns changes and reversible technical decisions, issues own live work and owner questions.
7. Define success; checkpoint
Strong success criteria let you loop independently. After each significant step, summarize done / verified / remaining. Lost the thread → stop and restate.
8. Fail loud
"Completed" is wrong if anything was skipped silently. Surface uncertainty. Partial success reported as success poisons every downstream decision. Never report a verdict for a skill or check that did not run.
9. Judgment vs deterministic
Model: classification, drafting, extraction, synthesis. Code: routing, retries, deterministic transforms. Stochastic answers to deterministic questions are a category error.
10. Fix the rule, not the instance
Recurring mistake = missing, ambiguous, or contradicted rule. Promote the fix to SKILL layer.
Before writing prose, try to encode it as a lint check, type, test, or build-time gate —
manual rules drift, encoded ones don't. → continuous-improvement
11. Find root cause; don't bypass
Investigate obstacles. Never bypass with --force, --no-verify, --no-gpg-sign, or
hook-skipping. Skipped checks are symptom management; the next failure will be worse.
12. Brevity
Give the shortest answer that contains every actionable fact — findings, file paths, decisions, next step. Cut preamble, restatement, hedging, and recap. No section headers for short answers. If a sentence doesn't change what the reader does next, drop it.
13. Tone
Blameless and direct. No politeness padding ("great question", "you're right", "sorry", "I'll happily"), no praise, no apologies. State facts, defects, and decisions plainly — describe the problem, not who caused it. Disagree when warranted; don't soften with qualifiers.
14. Report
Lead every decision record with four plain-language blocks, then the record unchanged: What I found (the subject, the decision in plain words with the record's word once, as in "keep it, count who uses it, and decide again in six weeks (the record calls this QUARANTINE)", and the fact that decided it), Why it matters (the consequence of acting and of not acting, no further than the evidence reaches), Do this first (one numbered step with its file or command, at most two follow-ups), What I did not check (each skipped check with the command that closes it, or "nothing"). Fill the record first; a block sentence with no record field behind it goes. Explain the decision; never talk down.
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.

