Skill validator

skill-validator — L5 quality gate (the second L5 gate)

skill-eval answers "does this skill change behavior?" — skill-validator answers "is this skill well-written?". Two different failure modes; a skill can move a number and still be bloated, vague, or scope-creeping. Both gate the framework.

Two halves

1. Deterministic budget scan (cheap, instant — no model):

node skills/l5-meta/skill-validator/budget.mjs            # scan every SKILL.md
node skills/l5-meta/skill-validator/budget.mjs --self-test

Budgets: description ≤ 1024 chars (agentskills.io spec); body soft-cap ~4000 chars, hard flag

6000 (a resident skill like ponytail pays this on every turn); > 200 lines flags sprawl.

2. Six-axis LLM judge (via subagents — one judge per skill): Each skill is scored 0-10 on —

1. clarity — unambiguous, well-structured, easy for a model to parse.
2. actionability — concrete executable steps/rules, not vague aspiration.
3. token-efficiency — concise for what it delivers; no bloat or repetition.
4. scope-discipline — one clear job; doesn't sprawl into other skills' territory.
5. directive-precision — the description says WHEN to use the skill, not just WHAT it does (a what-summary makes the model follow the description and skip the body).
6. novelty — actually changes behavior vs the model's default (not a no-op restatement).

Run the judge the same way skill-eval runs — through subagents when no API key is available (no claude -p dependency), or adapt to a direct-API runner. The judge model should be strong (opus); pin it and read scores as a trend, not an absolute bar (LLM-judge scores drift).

How to read

• Budget flag → mechanical fix (trim the body; resident skills are the priority).
• Axis < 6 → a real authoring problem; the top_issue names the single most important fix.
• A skill that scores high on skill-validator but Δ 0 on skill-eval is well-written but inert; one that moves a number but scores low here is effective but a maintenance liability. The bar is both.

This is the gate every new skill (and every borrow that becomes a skill) passes alongside skill-eval before shipping. Pairs with improve-loop (which records the results) and scripts/check-rule-copies.mjs (structural drift).