build-evaluate
Skill6 starsChanged 29 days agoMITView on GitHub
Verify a completed build against the PRD's anchor case and success criteria. Runs the acceptance test and anchor-case check (serial across drivers sharing a write root), fans out 2-3 read-only cold-read agents (fresh-eyes, anti-scope-creep, drift-vs-mirror). On FAIL, auto-invokes five-whys for root-cause. On PASS with novel patterns, invokes insight-promotion to codify learnings. Read-only against the built artifact, idempotent — safe to re-run. Use after build-execute reports success. Triggers — "evaluate the build", "did it work", "verify the build", "acceptance test", "is it actually done". Outputs `.claude/PRPs/{slug}/evaluate.md` with a binary verdict.
In this repository30 other files it gives its agents.
- adversarial-re-readSkill
- adversarial-reviewSkill
- build-estimateSkill
- build-executeSkill
- 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, 3 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 11
Verify a build against its own definition of "done."
- Distinctive: line 25
If anchor case or acceptance test is missing: refuse and tell the user to update the PRD/plan first.
- Distinctive: line 77
Do NOT edit the artifact under test.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 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 Evaluate
Verify a build against its own definition of "done."
Phase 1: PARSE + PARALLEL LOAD
Batch in one message — independent reads:
Read.claude/PRPs/{slug}/prd.md,plan.md,state.jsonRead.claude/PRPs/_profiles/{type}.md(from PRD frontmatter)Glob.claude/PRPs/*/evaluate.md— for insight-promotion later, find prior similar builds for pattern matching
Verify:
state.jsonstateissuccessorstopped(notrunning) — refuse to evaluate a build still iterating- PRD has a concrete anchor case
- Plan has an acceptance test defined
If anchor case or acceptance test is missing: refuse and tell the user to update the PRD/plan first.
Phase 2: LOAD PROFILE
The profile may define:
- A canonical acceptance-test invocation (e.g.,
python3 scripts/validate-plugin.py path/to/plugin) - A fixture set the build must run against
- A specific reviewer agent to run as a final cold-read pass
Phase 3+4: RUN ACCEPTANCE TEST + ANCHOR CASE
⚠ SERIAL BY DEFAULT. Parallelize only where the write roots are certified disjoint. "Independent" does not mean "measures a different thing" — it means no two concurrently-running drivers WRITE the same tree. A driver that invokes a generator and then reads back what it generated is NOT independent of another driver doing the same: they interleave, and each asserts against a tree the other is rewriting. The run still prints PASS/FAIL, so the damage is invisible — the result is UNMEASURED, not failed, and unmeasured results that look green are how a build ships on evidence it never had.
Before launching anything, resolve each driver the plan names to (a) the generator it invokes and (b) the root that generator writes. Drivers sharing a root run ONE AT A TIME, in the order the plan lists them. Only drivers over disjoint fixtures run together, in one message.
- Acceptance test from the plan — serial across drivers sharing a write root
- Anchor case check (run the artifact against the PRD's one concrete example, compare output to expected outcome)
⚠ A killed generator can leave its tree half-written. Prefer letting one finish over killing it. Where the generator rebuilds from inputs held OUTSIDE the tree it writes, one full serial run is both the repair and the measurement — no separate repair step is needed.
⚠ If you must stop a job, verify it is GONE — a signal is not a stop. A script whose
trap cleanup ... TERM handler does not call exit runs the handler and RESUMES; if that handler
deletes the scratch the run measures against, the run continues and keeps reporting rows it can no
longer have measured. Measured 2026-09-10 on run-mutations.sh. After any kill, ps for the
process; if it survived, SIGKILL is the honest signal where the script owns only disposable scratch.
Anchor case PASS = output matches the PRD's expected outcome (exact OR fuzzy per criteria the PRD specified — PRD is the contract).
Where the rest lives (read the file for the phase you are in)
| Phase / topic | Read |
| Phase 5 cold-read fan-out (read-only rule, three named agents) and Phase 5.5 synthesize | references/cold-read-and-synthesis.md |
| Phase 6 report template, two verdicts, Phase 6.5 branch on PASS / FAIL | references/report-and-verdict-branch.md |
| Phase 7 user verdict gate and final Output block | references/verdict-gate-and-output.md |
Guidelines
- This skill is the only one authorized to mark a build
complete. Execute reports success; evaluate confirms it. - The anchor case is non-negotiable. A build that passes generic criteria but fails the anchor case is FAIL.
- Idempotent — re-running is always safe and recommended after any change to the artifact.
- Do NOT edit the artifact under test. Only the report and the PRD's status.
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.

