Skill Validator — Ralph Loop

A continuous-iteration eval loop for skills. Geoffrey Huntley-style: run the skill, evaluate, fix, re-run until pass criteria are met or the budget is spent. The loop fails closed, not open — if the budget exhausts before pass, the skill is reported as failing with a numbered list of unmet criteria.

When to use

• After authoring a new skill, before promoting it for general use
• After editing an existing skill's prompt or pre-flight checklist
• When a skill behaves correctly in chat but fails in scripted runs (or vice versa)
• When promoting a skill from one project to another
• When validating that a skill works against a specific target repo

When NOT to use

• The skill is purely conversational with no measurable output — there's nothing to score
• The target repo is unavailable or empty — there's nothing to run against
• The skill is already covered by deterministic unit tests — those are cheaper and stricter

Inputs

• SKILL_PATH — path to the skill being validated (e.g., .claude/skills/quality-review/SKILL.md)
• TARGET_REPO — path to the repo the skill will run against
• PASS_CRITERIA — file or inline list of required behaviors. Each criterion is a yes/no the loop can score. Format: <id>: <description> (one per line)
• BUDGET — max iterations (default 5) and max minutes (default 30). Hard cap; the loop fails if either is hit.

The Loop

iteration = 0
while iteration < BUDGET.max_iterations:
    iteration += 1

    # 1. Run skill on representative input from TARGET_REPO
    output = run_skill(SKILL_PATH, sample_input(TARGET_REPO))

    # 2. Score output against PASS_CRITERIA
    results = score(output, PASS_CRITERIA)
    if all(results.values()):
        return PASS(iteration)

    # 3. Identify which criteria failed and why
    failures = [(id, why) for id, ok in results.items() if not ok]

    # 4. Generate a targeted fix to SKILL.md
    patch = propose_fix(SKILL_PATH, failures)
    apply(patch)
    log(iteration, failures, patch)

return FAIL(iteration, last_failures)

Phase 1 — Setup (one-shot, before the loop)

1. Read SKILL_PATH end-to-end. Note the skill's stated triggers, inputs, and outputs.
2. Read TARGET_REPO/CLAUDE.md (if present) to understand project conventions.
3. Pick 2–3 representative inputs from the target repo:
   • For a code-review skill: 2–3 files representative of the codebase (one per language / layer)
   • For a docs skill: 2–3 documents of the type the skill targets
   • For an analysis skill: 2–3 inputs spanning easy / moderate / hard
4. Read PASS_CRITERIA and confirm every criterion is binary (yes/no), not vibes-based.
5. Take a baseline measurement: run the skill once against each input, record outputs, score them. This is iteration 0.

If iteration 0 passes everything, the skill is valid as-shipped. Report and exit.

Phase 2 — Iterate

Each iteration:

1. Run the skill against each representative input (2–3 runs, parallel if independent)
2. Score outputs against PASS_CRITERIA. Each criterion: PASS / FAIL / DEGRADED
3. Diff results against the previous iteration. Did the last fix introduce a regression?
4. Diagnose the highest-priority failure. Is it a prompt gap, a missing pre-flight check, an under-specified output format, or a fundamental mismatch with the target repo?
5. Patch the SKILL.md with the smallest change that addresses the diagnosis. Append the rationale to the audit log, NOT just the diff.
6. Log to .skill-validator/<skill-name>/iteration-<N>.json:
   • inputs (paths or hashes)
   • outputs (verbatim or path)
   • scores (criterion-by-criterion)
   • diagnosis
   • patch (unified diff applied to SKILL.md)
7. Check budget — iterations remaining, minutes remaining. If exhausted, exit FAIL.

Phase 3 — Exit

PASS — every criterion green for two consecutive iterations against every representative input. Two consecutive is the stability check; one good run could be a fluke.

FAIL (budget exhausted) — emit:

• Final SKILL.md state (with all patches applied)
• A diff vs. the original
• A numbered list of criteria that never passed, with the last failure mode for each
• A recommendation: human review, scope reduction, or skill rewrite

FAIL (regression detected) — if a patch makes a previously-passing criterion fail, revert the patch, log the regression, and continue with the next-highest-priority unaddressed failure. If no clean patch path remains, exit FAIL.

Pass criteria — examples by skill type

• quality-review — produces 3 numbered findings tables (one per agent), at least one Critical or High finding when reviewing intentionally-flawed input, no fabricated file references
• five-whys — reaches a structural root cause within 5 iterations on the seeded test failure; ends with an explicit governance update recommendation
• writing-quality — flags 100% of seeded P0 AI-isms in detect mode; rewrites pass a re-detect with zero P0 hits in rewrite mode

Rules and constraints

• Read-only against TARGET_REPO by default. Never modify files in the target repo unless --write-target is passed and the user has explicitly approved it. The validator's job is to fix the SKILL, not the target.
• Audit every iteration. No silent passes. The log file is the artifact, not just the final result.
• Bound the budget. A loop with no budget is not a loop, it's a runaway. Default 5 iterations / 30 minutes is generous; lower it for cheap skills.
• Two-strike rule on patches. If two consecutive iterations apply patches and neither moves the score, stop and ask the user. Continued patching is bias-confirmation.
• Don't optimize for the criteria. The criteria are the proxy, not the goal. If the skill passes the criteria but obviously gets the wrong answer, the criteria need rewriting, not the skill.
• Failure is the success signal for the validator. A validator that always passes is not measuring anything.

Outputs

• .skill-validator/<skill-name>/iteration-<N>.json — per-iteration audit log
• .skill-validator/<skill-name>/SUMMARY.md — final pass/fail report
• A diff of SKILL.md showing what changed across the loop

Agent integration

• debugger — invoked on the first same-failure-mode iteration repeat. Reproduces the skill failure deterministically and narrows the cause before patching.
• error-coordinator — invoked when the validator runs against multiple representative inputs and 2+ fail in the same iteration. Correlates symptoms to find a shared cause in the skill (not in the inputs).
• code-reviewer — invoked on the SKILL.md patch before it is applied each iteration. Catches sloppy edits — broken frontmatter, accidental scope drift, removed pre-flight items — that would make the next iteration noisy. A Critical finding from code-reviewer reverts the patch and re-diagnoses.