| name | skill-evaluation |
| description | Evaluate whether a the reference deployment skill actually works — blind runner + separate judge against a bar, rubric scoring, pass-rate on re-runs, and defect diagnosis. Use when asked to evaluate a skill, eval skills, run a blind eval, score against a rubric, test a skill, audit skill quality, grade a skill, review the skill library, or before promoting a new skill to the company master prompt or the whole swarm. Not for writing skills (use skill-authoring-standards) or deciding which to create (use suggest-skills). |
| owner | platform |
| created | "2026-07-05T00:00:00.000Z" |
| created_by | claude-fable-5 |
| sources | adapted from dzhng/skills (eval-skills, MIT) + the reference deployment local skill-eval-harness criteria |
Skill Evaluation
A skill that has never been evaluated is a hypothesis. This skill turns "looks good" into evidence: a fresh runner executes the skill on a realistic case, a separate judge grades the artifact against a bar, and misses get diagnosed before anything is rewritten. The builder of a skill is never its only judge.
When to invoke
- A new skill was just written (minimum: static rubric, ideally one blind run).
- Before a skill is referenced in the company master prompt or propagated swarm-wide.
- Quarterly library audit, or when an agent visibly misused/ignored a skill in the wild.
- After editing a skill to fix a failure — re-test all its cases, not just the one that failed.
Not for: grading one-off outputs (use review-output if available locally, or a judge subagent), or evaluating loops (use loop-library audit criteria).
Two evaluation modes — pick deliberately
- Judgment mode (default): the judge holds the artifact to a bar ("a zero-context agent following this would not lose data") — not to an expected output. Most the reference deployment skills encode judgment; grading them against a fixed checklist degenerates into conformance theater that stops testing the judgment.
- Conformance mode (only for fragile/mechanical skills): exact expected outputs/commands, byte-level where it matters (e.g. rollout steps, API auth headers).
Protocol
Phase 1 — Static rubric (~10 min, no execution)
Score the SKILL.md itself, 0/1 per row:
| # | Criterion |
|---|
| 1 | Frontmatter complete; description states what + literal trigger keywords |
| 2 | Trigger test: would skill_search with the obvious keyword actually surface it? (run it) |
| 3 | When-NOT-to-use present with sibling pointer |
| 4 | Protocol steps end in observable completion criteria |
| 5 | Output contract explicit (what the agent must return/produce) |
| 6 | Anti-patterns/gotchas present and earned (not generic filler) |
| 7 | Verification section demands evidence, not vibes |
| 8 | Provenance/maintenance: volatile facts date-stamped with re-verification commands |
| 9 | No unverified claims, no invented paths/commands (spot-check 3) |
| 10 | Size discipline: body ≤ ~2,500 words, references split out |
| 11 | No duplication: no fact owned by another doc restated here |
| 12 | Failure-mode scan: none of the taxonomy modes from skill-authoring-standards present |
≤7/12 → the skill needs work before any dynamic eval. Record the score in the eval report.
Phase 2 — Blind run (the real test)
- Write a golden case: a realistic input/task the skill should handle, WITHOUT hinting at the skill's internal steps. Include what a strong result must contain (the bar), kept hidden from the runner.
- Runner: a fresh agent (subagent, or a cheap-tier session — Sonnet-class, since that's the real swarm consumer) gets ONLY the case + access to the skill. No conversation history, no author context.
- Judge: a different fresh agent gets the runner's artifact + the bar. Prompt it to try to prove the output does NOT meet the bar. The judge never sees the runner's reasoning, only the artifact.
- Nondeterminism: borderline results get 2-3 re-runs; report a pass rate, not a verdict. A skill that passes 1 of 3 is not fixed.
Phase 3 — Diagnose before editing
For each miss, decide: skill defect (the skill misled or under-specified) vs bad case (the case demanded something outside the skill's job — fix the case or the bar, not the skill). Editing a skill to chase a wrong bar makes it worse. Map defects to the failure-mode taxonomy, fix, then re-run all cases.
Phase 4 — Report
Write findings to knowledge/platform/strategy/skill-eval-<scope>-<date>.md (or the skill's own Cambios entry for a single-skill eval): rubric score, cases run, pass rate, defects found with taxonomy label, fixes applied, what remains open.
Automated batch option (Mac-local, verified 2026-07-05)
the founder's Mac has a batch harness: ~/.claude/skills/skill-eval-harness/ → cd /workspace && make skill-eval-harness → report in outputs/ai-native-ops/YYYY-MM-DD/skill-eval-harness/. It runs static criteria (rubric rows 1-7 approximately) over local skills. It only exists on the founder's Mac — swarm agents apply Phase 1 manually. Porting it to the VPS is a future candidate (Codex task).
Guardrails
| ❌ | ✅ |
|---|
| Builder grades own skill and ships | Separate judge, fresh context, tries to refute |
| Judge gets the expected output | Judge gets the bar; expected outputs only in conformance mode |
| One lucky pass = validated | Pass rate over re-runs on borderline cases |
| Failing eval → rewrite skill immediately | Diagnose defect-vs-bad-case first |
| Eval finds 12 weak skills → silently rewrite all | the change-approval rule: report the list, the founder decides ✅/❌ per skill |
| Runner is a frontier model | Runner matches the real consumer tier (Sonnet/GPT-5.5 class) |
Verification
Done when: rubric scores recorded, at least one blind run per new skill with judge verdict, report path exists and is readable via brain_read.
Cross-references
- Authoring standard the rubric enforces: skill-authoring-standards
- What to create/improve: suggest-skills
- Adversarial judging for high-stakes outputs: internal
punta-de-flecha skill
- Builder != judge doctrine: internal
fable-prompting skill, section 4
The internal cross-referenced skills are named for context but are not included in this public package.
Provenance and maintenance
- Mac harness path + make target: verified 2026-07-05 by reading
~/.claude/skills/skill-eval-harness/SKILL.md. Re-verify: ls /workspace/Makefile on the Mac.
- Swarm consumer tiers (Sonnet/GPT-5.5): re-verify
brain_read("knowledge/platform/agents/model-routing-real-2026-07-04.md").
Cambios
| Fecha | Cambio | Por |
|---|
| 2026-07-05 | Created — dzhng eval-skills blind-run/judge/pass-rate mechanics + the reference deployment local harness rubric, adapted to Brain + swarm tiers. | Claude Code (Fable 5) |