Stats
Actions
Available In
Tags
Schliff
Your AI instructions are silently degrading. Schliff catches it.
Deterministic quality scoring for CLAUDE.md, SKILL.md, .cursorrules, AGENTS.md, and system prompts. No LLM, no API key — same input, same score. Python 3.9+, zero core dependencies (optional schliff[evolve] adds litellm for the evolution loop).
pip install schliff
schliff score path/to/SKILL.md
$ schliff score demo/bad-skill/SKILL.md
schliff v7.2.0
structure ███████░░░ 70/100 fair
efficiency ████░░░░░░ 35/100 poor
composability ███░░░░░░░ 30/100 poor
clarity █████████░ 90/100 great
Structural Score ███████████░░░░░░░░░ 53.8/100 [D]
⚠ 4/8 dimensions measured (weight coverage: 40%). Unmeasured: triggers, quality, edges
→ 13 deterministic fixes available. Run `/schliff:auto` in Claude Code to apply.
Tokens: 378 / 1,000 (ok)
A real optimization
@wan-huiyan ran schliff on the 1,331-line SKILL.md for agent-review-panel, a multi-agent code-review skill. Two optimization rounds later: 340 lines, 75% fewer tokens, structure 65 → 100, composability 56 → 91. A/B tested on a 1,132-line document — identical review quality with a quarter of the tokens.
| Skill | Score | Rounds | Author |
|---|---|---|---|
| agent-review-panel | 75 [C] → 85.6 [A] | 2 | @wan-huiyan |
| shieldclaw (OpenClaw) | 68 [C] → 94.6 [A] | 1 | @Zandereins |
| demo bad-skill | 54 [D] → 98.3 [S] | 18 auto | @Zandereins |
Score yours: schliff score path/to/SKILL.md — share what you find
Seen in the wild
A root CLAUDE.md written for modelcontextprotocol/servers (Anthropic's official MCP reference repo) merged to main on April 17th, 2026. Running schliff on it returned 59.2/100 at 40% weight coverage — a useful measurement of where the file actually needed work and where the scorer was structurally unfair for a project-root document. Full walkthrough →
What the data says
We scored 120 public instruction files across 60 source repos. Mean grade: D. 59% below C. Adding one companion eval suite lifts the mean +22 points.
- Composability is the real weak spot — mean 30.4/100. Files tell agents what to do, rarely where to stop or hand off
- No companion eval suite in the corpus — verified 0/60 source repos ship an
eval-suite.json,evals/, or any test artifact. Three dimensions stay unmeasured, locking 45% of the score - Hedging dilutes intent — efficiency averages 52.8/100. "You might want to consider" is noise
- Format alone doesn't save you — AGENTS.md averages 64.8, SKILL.md 55.4. Skipping frontmatter costs ~15 points regardless of format
Read the full report → · Reproduce it
What Schliff Catches
| Dimension | Weight | What it catches |
|---|---|---|
| structure | 15% | Missing frontmatter, empty headers, no examples, dead content |
| triggers | 20% | Eval-suite trigger accuracy, false positives, missed activations |
| quality | 20% | Thin assertions, missing feature coverage, low coherence |
| edges | 15% | No edge cases defined, missing categories (invalid, scale, unicode) |
| efficiency | 10% | Hedging, filler words, repetition, low signal-to-noise |
| composability | 10% | Missing scope boundaries, no error behavior, no handoff points |
| clarity | 5% | Contradictions, vague references, ambiguous instructions |
| security | 5% | (opt-in) Hardcoded secrets, unsafe commands, exposed credentials |
Grades: S (≥95) · A (≥85) · B (≥75) · C (≥65) · D (≥50) · E (≥35) · F (<35). Full methodology: docs/SCORING.md
