What

Schliff is a deterministic quality scorer for AI instruction files — CLAUDE.md, SKILL.md, .cursorrules, AGENTS.md, and system prompts. It evaluates files across 7 weighted dimensions (structure, triggers, quality, edges, efficiency, composability, clarity) using static analysis with no LLM in the scoring path. schliff suggest returns ranked fixes with estimated point impact; schliff verify --regression gates CI on score drops. An optional schliff evolve loop (pip install schliff[evolve]) applies deterministic patches first and falls back to an LLM for structural changes rules can't make.

Why

Instruction files degrade silently — scope dissolves, hedging creeps in, edge cases stay undefined. We scored 120 public instruction files across 60 source repos: mean grade D, 59% below C, and zero ship a companion eval suite. Adding one lifts the mean composite +22 points. The deterministic scorer makes regressions visible and gateable the same way test-coverage regressions are.

Schliff complements the existing entries in Development & Code Tools: Skill Creator builds v1 skills, Skill Seekers converts docs to skills, Schliff scores and improves them — the third leg of the skill lifecycle.

Evidence / Tested

• Live on PyPI: pip install schliff — Python 3.9+ stdlib only (zero core deps); optional schliff[evolve] adds litellm
• 1,007 unit tests + 89 integration tests, CI green on Python 3.9 and 3.12
• Full 120-file quality report with methodology + raw scores
• External case study: @wan-huiyan used schliff to reduce agent-review-panel's SKILL.md from 1,331 to 340 lines (75% fewer tokens, A/B-validated identical quality) — documented at HOW_WE_BUILT_THIS.md
• Determinism is testable directly: schliff score <file> produces identical output on every run

Changes in this PR

One-line addition to README.md under Development & Code Tools. Refreshed 2026-04-17 to reflect the current four-format coverage (CLAUDE.md, SKILL.md, .cursorrules, AGENTS.md) and CI-gating surface (schliff verify --regression) — prior text mentioned only SKILL.md.

Attribution

Built by @Zandereins (Franz Paul) with Claude Code.