build-execute
Skill6 starsChanged 28 days agoMITView on GitHub
Bounded execution loop for a validated build plan. Hard caps on iterations, edits, and (where applicable) spend prevent runaway loops. Each iteration runs a CONDITIONAL adversarial review (only when the change touches a contract, a pinned constant, a refusal limb, or an unmutated path), applies the edit, measures via parallel fixture runs, and auto-invokes five-whys on stagnation or regression. Calls writing-quality on prose edits. Stops on success, on cap hit, on stagnation, or on regression. Use after Stage 3 is dispositioned (PASS, SKIPPED or WAIVED - recorded either way), before build-evaluate. Triggers — "execute the build", "build it", "iterate the X", "implement the plan". Refuses to start if the probes mismatch, the plan-soundness gate fails, the Stage 3 disposition is unrecorded, or profile preconditions are unmet.
In this repository30 other files it gives its agents.
- adversarial-re-readSkill
- adversarial-reviewSkill
- build-estimateSkill
- build-evaluateSkill
- build-planSkill
- build-polishSkill
- build-prdSkill
- build-probeSkill
- build-validateSkill
- excel-connector-governanceSkill
- five-whysSkill
- insight-builderSkill
- insight-crystallizerSkill
- insight-promotionSkill
- multi-sessionSkill
- notion-governanceSkill
- oodaSkill
- prd-writerSkill
- promote-to-toolingSkill
- quality-reviewSkill
- ralph-implementSkill
- ralph-loopSkill
- search-governanceSkill
- session-outputSkill
- skill-validatorSkill
- task-planSkill
- verifySkill
- voice-profile-testSkill
- why-diagnosticSkill
- writing-qualitySkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 2 are distinctive, and 1 is 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.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 21
No prior .claude/PRPs/{slug}/state.json is in state: running (refuse double-runs)
- Distinctive: line 23
If any precondition fails: stop, list the failures, do nothing else.
- Distinctive: line 50
After preconditions PASS but BEFORE iteration 1 starts, present the run plan and ask for confirmation:
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill 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
Build Execute
Bounded execution loop. The scheduler and budgeter; per-iteration work is delegated.
Hard preconditions (REFUSE if unmet)
.claude/PRPs/{slug}/prd.mdexists.claude/PRPs/{slug}/plan.mdexistsbash .claude/PRPs/{slug}/probes/run.shre-run in THIS session: 0 MISMATCH, 0 source changedcheck-plan-soundness.py {slug}/plan.md --prd {slug}/prd.mdexits 0- Stage 3 disposition is RECORDED in
state.jsonasvalidate_regimeand is one of the three below - Profile's hard preconditions (defined in the profile) are met
- No prior
.claude/PRPs/{slug}/state.jsonis instate: running(refuse double-runs)
If any precondition fails: stop, list the failures, do nothing else.
The three Stage 3 dispositions
Premises get reviewed once — confirmed, contradicted, or decided. None of those improves by re-reviewing, so "validate again until PASS" is not a thing. Record exactly one:
validate_regime | When | Must also record |
| PASS | Panel ran, frame held | Date, which premises were contested |
| SKIPPED | Nothing contested — no new module, no schema change, no reversal of what a field MEANS. The common case on a running build | Suite green, mutations FULL PASS, driver written RED first |
| WAIVED | Panel ran, FAILED, owner overrode on the record | Who, when, why, substituted gate, and what_is_NOT_claimed |
⚠ SKIPPED needs a mutation-proven suite. Green but unmutated is unfalsifiable — that reverts the target to greenfield, which takes the panel. So does a schema change or a new module, whatever the surrounding suite proves.
⚠ Three FAILs on one target is a process defect (Hard Rule 9). Re-scope; don't convert a repeated FAIL into a waiver.
Hard caps (defaults — overridable per invocation but not per iteration)
- Max iterations: 3 (override
--max-iterations N) - Max edits per iteration: defined by profile (e.g., skill = 1 file, plugin = N skills × 1 file each)
- Max spend: profile-defined. Skip if profile has no spend (most internal builds have none)
- Stagnation: 2 consecutive iterations with no measurable progress on the acceptance test → stop automatically
Caps are enforced BEFORE each new iteration. They cannot be raised inside the loop.
Pre-iteration-1 gate (REVIEW GATE — skip with --autonomous)
After preconditions PASS but BEFORE iteration 1 starts, present the run plan and ask for confirmation:
Print:
- Shape: {single-shot / iterative-bounded / pipeline / wrapper}
- Max iterations: {N}
- Max edits per iteration: {profile-defined}
- Files that will be touched: {list from plan Step-by-step tasks}
- Acceptance bar: {from PRD}
- Estimated effort: {from plan}
AskUserQuestion:
Question: "Plan ready to execute for {slug}. Begin iteration 1?"
Options:
- "Yes, begin" → start iteration 1
- "Yes, but dry-run first" → re-invoke with --dry-run, present what would happen, ask again
- "Adjust caps" → ask follow-up for new --max-iterations / --max-spend, then begin
- "Stop" → exit cleanly, leave state.json with state: aborted-pre-iterationWhy: the first edit commits the direction. If the plan says "rewrite SKILL.md prose for clarity" but the user realizes the actual issue is the trigger phrasings, this is the cheapest moment to course-correct.
Skip this gate if --autonomous was passed (e.g., for scheduled/batch runs).
Where the rest lives (read before running the matching phase)
| When you are about to... | Read |
| Run an iteration (Steps 1-7: read state, decide, baseline, adversarial review, writing pass, apply edit, measure, continue/stop) | references/iteration-steps.md |
| Write the iteration report (Step 8), update state, produce Output or the session report, run dry-run mode, or check the Guidelines | references/report-state-output.md |
The hard preconditions, hard caps and review gate above apply before and during every iteration. Do not start an iteration without reading references/iteration-steps.md.
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.

