snes-fit
Fleet-wide QA skill that audits any other skill against structural, adversarial, scale, composition, security, threat-intel, type, and design dimensions. Plugin architecture lets you add dimensions as separate dirs — no core code changes. Ship only when decay rule hits 2 consecutive zero-finding rounds. Compatible with Anthropic skill-creator eval schemas. NOT for: creating new skills (use snes-builder). NOT for: running skills (use the skill directly). NOT for: continuous background monitoring (use scheduled-tasks).
Resources
11Install
npx skillscat add dimmmak/snes-fit Install via the SkillsCat registry.
.snes-fit — Fleet-Wide QA Skill
Purpose
Runs 8 pure-Python dimensions against any skill in the fleet. Binary PASS/FAIL/UNKNOWN verdicts; aggregated into a letter grade; ship gated by the decay stopping rule (≤1 cosmetic + 0 structural/major findings for 2 consecutive rounds).
Phase 1 is stdlib-only. No Claude API key required. Phase 2 adds LLM-backed dimensions (executor + judge).
snes-fit is the structural-conformance auditor of the fleet — it grades skills against SPEC.md. Use .forensic for editorial / quality critique on artifacts; use snes-fit for spec-compliance.
When to trigger
Fires when the user invokes any of:
.snes-fit audit --skill <name>/.snes-fit audit --all.snes-fit design-audit --skill <name>.snes-fit create --skill <name>/.snes-fit improve/.snes-fit benchmark- Natural-language phrasings: "snes-fit this", "audit skill X for spec compliance", "run the fleet QA on X", "score skill X", "grade skill X against spec", "is X ship-ready?", "check fleet conformance"
- Any request to verify a skill's frontmatter, required sections, forbidden patterns, or canonical subdirectories against SPEC.md
- Self-audit:
.snes-fit audit --skill snes-fit(the auditor must audit itself; seedimensions/00_self/)
When NOT to trigger
- Editorial / logical / quality critique on an artifact — that's
.forensic. snes-fit is structural; forensic is mechanism-and-quote. - Creating a new skill from scratch — that's
snes-builder. snes-fit grades; it does not author. - Running a skill — that's the skill itself. snes-fit is a meta-tool; calling it does not invoke the audited skill.
- Continuous background monitoring — use
scheduled-tasksto trigger snes-fit on a cadence; snes-fit itself is one-shot. - Cross-skill code execution / orchestration — that's
mewtwo. snes-fit reads source files; it does not run them. - Live model comparisons or A/B prompt evaluation — out of scope (NON_GOALS.md #6).
Anti-patterns
- Audit drift — running snes-fit, finding zero issues, declaring victory. The decay rule requires 2 consecutive zero-finding rounds, not one. Single-pass green is not ship-ready.
- Self-audit skip — auditing every skill except snes-fit itself. If the auditor doesn't pass its own bar, every grade it emits is suspect. Run
audit --skill snes-fitbefore any fleet sweep. - Severity inflation — tagging every minor finding as "critical" to force attention. Severity tiers are critical / major / minor / cosmetic per SPEC.md:108-115. Do not invent tiers.
- Score gaming — exempting unfavorable findings via private skip-list. SPEC.md is the single source of truth; any exemption must be a SPEC.md edit + version bump, not a config-file workaround.
- Compatibility theater — claiming "Anthropic skill-creator eval schemas" compatibility without a pinned schema version + test. Pin the version; commit the test.
- Composability claim without bilateral check — declaring
composable_with: mewtwoin frontmatter without verifying mewtwo's contract accepts snes-fit as a caller. Mewtwo refuses skills lacking contract declarations.
Exit conditions
- Audit complete — scorecard emitted to
reports/<date>-<skill>.md, findings tovault/<skill>/findings.jsonl, decay state updated. Skill done for this round. - Decay rule satisfied — 2 consecutive rounds with ≤1 cosmetic + 0 structural/major. Ship gate opens. Skill done.
- Critical finding — block the round, surface to user, exit. Critical = broken or dangerous (per SPEC.md:111). Cannot ship past critical.
- Self-audit failure — if
audit --skill snes-fitproduces critical findings, snes-fit refuses to fleet-sweep until self-audit clears. The auditor must clear its own bar first.
Subcommands
| 🟣 Command | 🟣 Mode | 🟣 What it does | 🟣 Writes |
|---|---|---|---|
.snes-fit audit --skill <name> |
Eval | Runs all enabled dims, prints scorecard | vault/<skill>/findings.jsonl, reports/<date>-<skill>.md |
.snes-fit audit --all |
Eval | Fleet sweep — all skills under ~/Desktop/CLAUDE CODE/ |
per-skill vault + one summary report |
.snes-fit design-audit --skill <name> |
DesignAudit | Runs ONLY dimension 08 — fast structural/design check | vault/<skill>/findings.jsonl |
.snes-fit create --skill <name> |
Create | Scaffolds evals/<skill>/evals.json (Anthropic-compatible) |
evals/<skill>/evals.json |
.snes-fit improve --skill <name> |
Improve | Phase-1 stub; phase-2 LLM fix engine | — |
.snes-fit benchmark --skill <name> |
Benchmark | Rolling baseline; delta vs previous run | vault/<skill>/benchmark.jsonl |
4-mode lifecycle
| 🟣 Mode | 🟣 Input | 🟣 Output | 🟣 Purpose |
|---|---|---|---|
| Create | skill name | evals.json scaffold |
Seed an eval suite |
| Eval | skill + evals | scorecard + findings | Run once; grade it |
| Improve | findings | fix proposals (phase 2) | Close the loop |
| Benchmark | skill + history | delta report | Track drift over time |
Plus two always-on gates: Audit (all dims) and DesignAudit (dim 08 only).
Decay stopping rule
Per principle_stress_test_cadence — ship only when findings decay to ≤1 cosmetic + 0 structural for 2 consecutive rounds. Each round = one full audit pass with fixes in between. The decay tracker in scripts/lib/decay_tracker.py enforces this automatically; no bypass.
Plugin tree
Every dimension is its own dir under dimensions/. Drop in dimensions/NN_name/plugin.py with a DimensionPlugin subclass and it gets auto-discovered. No core code touched.
See ARCHITECTURE.md for the DimensionPlugin ABC, data flow, and phase roadmap. See SCHEMA.md for every file format. See NON_GOALS.md for what this skill will never do.