Turn prompt folders into falsifiable, versioned contracts your CI can enforce
skill 纪律层：SPEC 模板 + 证据流水线 + Prompt-as-Source 模式治理（吸收 skill-spec / agentsrc）
git clone https://github.com/chenhz01/skill-os| skill-os | 裸提示词文件夹 | |
|---|---|---|
| 契约 | SPEC 六节+版本化 | 无 |
| 可审计 | evidence 流水线 | 无 |
Topics: agent-skills · spec · ci · prompt-engineering · zero-dependency
Part of the Evidence-first Agents suite — tooling that makes AI agents accountable instead of just capable: answer-contract (output discipline) · skill-spec (spec discipline) · skill-scorecard (buyer-side grading) · memory-wiki (memory discipline). skill-os is the assembly line that wires all four into one pipeline.
The assembly line for evidence-first agent skills. Four single-file tools, one pipeline. Battle-tested on a 482-skill internal library (full sweeps, N/N coverage, no sampling).
| Stage | Tool | Repo | Question it answers |
|---|---|---|---|
| 0 contract | answer-contract | chenhz01/answer-contract | How should the agent talk? |
| 1–2 spec | skill-spec | chenhz01/skill-spec | How does the agent work, and what proves it ran? |
| 3 grade | skill-scorecard | chenhz01/skill-scorecard | Is this skill safe to install? |
| 4 memory | memory-wiki | chenhz01/memory-wiki | What did the agent learn, and is it still fresh? |
skill_os.py wires them into one line per skill:
contract -> lint -> evidence -> scorecard -> verdict -> memory
python tools/skill_os.py check <skill_dir> # full line on one skill
python tools/skill_os.py sweep <library> --md r.md --wiki wiki/ # full coverage N/N
python tools/skill_os.py run <skill_dir> --note "" # generate evidence run file
python tools/skill_os.py audit <evidence_file> # all boxes ticked?- BLOCKED — lint FAIL, or grade D, or any scorecard FAIL. Do not ship.
- NEEDS-EVIDENCE — structure fine, but no fully-ticked evidence run. Don't claim done.
- SHIP-READY — lint pass + latest evidence run fully ticked + grade A/B.
obra/superpowers — a popular real-world skill collection, audited with sweep on 2026-09-16 (snapshot b36e0829c6d0, full coverage 14/14, no sampling):
| Check | Result |
|---|---|
| lint (structure) | 14/14 pass |
| evidence (audit runs) | 0/14 fully audited |
| scorecard grades | A=0 · B=12 · C=2 · D=0 |
| verdicts | NEEDS-EVIDENCE=14 · BLOCKED=0 |
Honest reading: superpowers is structurally clean — every skill lints, nothing blocked, nothing graded D. The two C grades (brainstorming, subagent-driven-development) bundle a scripts/ directory without a license file or SPEC. And none of the 14 carries per-run evidence — exactly the gap this assembly line exists to close: a clean-looking skill and an audited skill are not the same thing.
skill_os.py looks for the four tool files beside itself (sibling repos layout) or
dropped into its own tools/ directory. Zero dependencies, Python 3.10+ stdlib only.
- lint green = structure complete, not human-reviewed.
- Grade A = structurally trustworthy, not provably safe.
- Evidence audited = every box ticked with real artifacts; artifact quality still needs judgment.
- The sweep reads files only; it never executes skill scripts.
MIT