agentleFS
puedingpueding/houserules/skills/test/SKILL.md

test

Skill1 starsChanged 3 months agoMITView on GitHub

Empirically test whether proposed CLAUDE.md rules actually change behavior, before applying them. Runs each rule A/B against a real probe prompt in separate headless Claude Code processes, reports FIRES / PARTIAL / NO EFFECT / BACKFIRES from deterministic observables, and drafts sharper wording for the ones that do not land. Use when the user runs /houserules:test, asks to test or validate proposed rules, asks whether a rule actually works, or wants to verify a CLAUDE.md change before committing to it.

184 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use agent-context-isolation, agentica-infrastructure and 10 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • Other versions184 projects use a version of this.

    Word for word · 0

    None found.

    Edited · 0

    None found.

    Wrote their own version · 184

    1. react/react★ 253ksame name
    2. facebook/react★ 253ksame name
    3. apache/skywalking★ 25ksame name
    4. livewire/livewire★ 24ksame name
    5. NVIDIA/TensorRT★ 13ksame name
    6. FalkorDB/FalkorDB★ 11ksame name
    7. FastLED/FastLED★ 7.5ksame name
    8. fleetdm/fleet★ 7.2ksame name
    9. parcadei/Continuous-Claude★ 3.9ksame name
    10. parcadei/Continuous-Claude-v2★ 3.9ksame name

    174 more not listed

    How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.

  • In this repository3 other files it gives its agents.

The file

We split this skill into the rules it gives an agent

4 rules. 0 are common, 3 are distinctive, and 1 is very specific to this project.

  • 0

    Common

    Most of the 158 edited versions we read keep these.

  • 3

    Distinctive

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Never more."* produced ~200 words with headers and bullets.

    line 13
  2. Very specific:

    If ${CLAUDE_PLUGIN_ROOT} is unset — running the skill from a clone rather than a plugin install — use the path to this repo's scripts/ …

    line 54
  3. Distinctive:

    Do not spend the user's time rediscovering this.

    line 78
  4. Distinctive:

    Do not loop indefinitely; report both attempts and hand the decision back.

    line 105

“Common” is measured: kept by at least 6 in 10 of the 158 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

houserules:test

A rule is not right or wrong in isolation — it competes with everything already in the file, and a well-worded rule can lose. You cannot see that by reading.

Measured on a real machine: appending "Answer in exactly one short sentence. Never more." produced ~200 words with headers and bullets. It lost to the style rules already present — gloss-in-parens, everyday-analogy, concrete-examples — which are visible firing in the output. The same rule substituted in place of those rules produced one sentence. Same rule, same probe, same model.

That is the failure mode behind learn's best finding, "the rule exists and is not working." This skill catches it before the rule is applied instead of weeks later.

Probes run as separate processes, never subagents

A subagent shares this session's harness and cannot load a different CLAUDE.md, so it cannot vary the thing under test. Every probe is a claude -p subprocess with its own harness resolution. They parallelize as background jobs, so this costs wall-clock, not correctness.

Each probe is a full billable Claude Code invocation. A 7-row table at n=3 is ~42 runs. Say the number before starting and let the user cut it.

Phase 1 — Build a probe per rule

The best probe is the correction that produced the rule. learn rows carry verbatim prompts the user actually typed before complaining. Replay one: if the rule is real, that same prompt stops producing the thing they complained about.

Rule sourceProbe
learn rowthe verbatim prompt from its evidence — always prefer this
migrate rowsynthesize one that would exercise the behavior the guide describes
Rule about not editinga prompt that tempts an edit; observable is whether files changed

A probe must be able to fail. "Explain X" cannot test a conciseness rule if the control answer was already short. Check the control first.

Phase 2 — Run

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/probe.py" \
  --probe probe.txt --new rule.txt --mode add -n 3 --model <target model>

If ${CLAUDE_PLUGIN_ROOT} is unset — running the skill from a clone rather than a plugin install — use the path to this repo's scripts/ directory instead. The scripts are standalone and dependency-free.

Use the model that will actually read the file. A haiku result does not transfer to opus; the whole question is how this rule behaves against that model's defaults.

Modes are different experiments — do not conflate them.

  • --mode add — the rule is new. Arm B appends it, arm A does not, the global file is identical in both and cancels out. Nothing is written. Always works.
  • --mode rewrite — the rule replaces existing text. Appending would test "new on top of old", the wrong question, and reliably reports NO EFFECT. A true substitution needs --allow-swap, which briefly stands a variant in place of the real CLAUDE.md (backed up, restored in a finally, refuses to start if a previous run left one behind).

    Ask before using --allow-swap. It touches the live file. Without it the rewrite is tested as an add and reported INCONCLUSIVE — never present that as a verdict.

A shadow config dir via CLAUDE_CONFIG_DIR does not work: on a live machine the on-disk credentials are stale and a separate config dir cannot refresh them. Do not spend the user's time rediscovering this.

Phase 3 — Verdict

Decide from the deterministic observables first — words, lines, bullets, headers, code_fences, and for edit-scope rules whether files changed. Reach for a judgment call on the text only when the numbers do not settle it.

VerdictMeaning
FIRESObservables move the intended way, consistently across runs
PARTIALSome clauses land, others do not — the most useful verdict
NO EFFECTNo consistent movement; the rule is being outcompeted
BACKFIRESMovement in the wrong direction
INCONCLUSIVEArms errored, or a rewrite was tested as an add

PARTIAL is the verdict to look for. A measured example: "Answer in a single short paragraph. No bullet lists, no headers, no code blocks." took bullets from 2 to 0 (−100%) and lines from 17 to 11 (−35%), while words went up 9%. The formatting clause landed; the brevity clause did not. One rule, one clause to fix.

Report n and the per-run spread. 2/3 is not 3/3, and a single run is indicative, not conclusive — say so rather than implying certainty you do not have.

Phase 4 — Revise what did not land

For NO EFFECT, BACKFIRES, or PARTIAL, draft a sharper version and retest once. Do not loop indefinitely; report both attempts and hand the decision back.

What usually makes a rule lose, in order:

  1. Outcompeted. Another rule in the file pulls the other way and is more specific or more vivid. Name the competing line. Often the fix belongs there, not in the new rule.
  2. Abstract. "Be concise" states a value with no observable. "Lead with the answer in one sentence, detail after" states a behavior.
  3. Prohibition without replacement. "Don't over-explain" leaves nothing to do instead. Positive instructions beat negative ones — say what to produce.
  4. Buried. Correct wording, wrong place, nothing keys it to a trigger.

Quote the competing line when you find one. "This lost to L19-21" is actionable; "try rewording" is not.

Phase 5 — Report

| # | Rule                   | Probe source      | n   | v1        | v2 (revised) |
| 1 | Default to high-level  | your 07-26 "tldr" | 3/3 | FIRES     | —            |
| 2 | Don't edit on analyze  | your 07-29 prompt | 0/3 | NO EFFECT | FIRES 3/3    |
| 3 | Deliverable length     | synthesized       | 2/3 | PARTIAL   | FIRES 3/3    |

Then per row: the probe, the observable deltas, the verdict, and — where revised — both wordings with what changed and why.

Hand back to whichever command produced the rules. This skill never writes to CLAUDE.md; it tells the user which rules are worth applying.

Anti-goals

  • Do not use subagents for probes. They cannot vary the CLAUDE.md under test.
  • Do not test on a cheaper model than the one that will read the file.
  • Do not report a rewrite-tested-as-add as a verdict. It is INCONCLUSIVE.
  • Do not use --allow-swap without asking. It touches the live file.
  • Do not claim a verdict from n=1. Say it is indicative.
  • Do not loop revisions. One retest, then hand it back.
  • Do not write to CLAUDE.md. Reporting only.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.