skill-update

Skill Update

Generate or refine Claude Code skills following Anthropic best practices.

Hard caps (enforced by scripts/skill_linter.py)

• SKILL.md ≤ 150 lines
• reference/*.md ≤ 200 lines (reference/scenarios/*.md ≤ 400 lines)
• README.md ≤ 100 lines
• Every SKILL.md has YAML frontmatter with name + description
• No DO NOT / MUST NOT / NEVER outside an ## Anti-Patterns section
• No challenge-specific identifiers (machine names, lab IDs, lab IPs, preserved flags)
• Every Markdown link resolves to an existing file
• Every reference file is linked from at least one other file (no orphans)

Principles

• Brevity first. Every file short, simple, human-readable. Challenge every token.
• Progressive disclosure. SKILL.md navigates; reference/ holds detail; reference/scenarios/ holds concrete exploit flows.
• Separation of concern. SKILL.md = WHAT + when. reference/role-*.md = HOW agents behave when spawned.
• Single canonical home for any cross-cutting rule (output discipline, credential loading, brute-force, etc.). Other files reference, never restate.

File structure

skills/<skill-name>/
├── SKILL.md           # ≤150 lines, YAML + navigation
├── reference/
│   ├── *-principles.md  # ≤150 lines (decision tree)
│   ├── INDEX.md
│   ├── *.md             # patterns, ≤200 lines
│   └── scenarios/
│       └── <category>/
│           └── *.md     # ≤400 lines, self-contained
└── README.md          # optional, ≤100 lines

SKILL.md template

---
name: <skill-name>
description: What it does AND when to use. Include trigger phrases.
---

# <Skill Name>

<one-paragraph scope>

## When to use

- <bullet>

## Workflow / Quick start

<≤30 lines>

## References

- [reference/...](reference/...)

## Anti-Patterns

- <when negative framing is genuinely needed, put it here>

Run it

The procedure is a workflow — every step is code-enforced, so none can be skipped.

Workflow('skill-update', { output_dir: 'projects/<engagement>' })   // harvest an engagement
Workflow('skill-update', { learnings: [{ text, technique_type }] }) // judge a known set
Workflow('skill-update', { mode: 'audit' })                         // read-only, writes nothing

Add dryRun: true to see the write plan without writing. Parent-orchestrator only.

Phases. Intake (baseline skill_linter.py --json) → Harvest (reframe learnings) → Judge (four gates + blind refuters) → Route (author the block) → Write (persist verbatim) → Sweep (confidentiality guard) → Verify (linter delta, not an absolute clean tree — the base carries pre-existing violations).

Every promote/reject/write decision is pure JS in .claude/workflows/lib/wf-helpers.mjs (promotionGate, capBudget, writeGate, lintDelta, skillUpdateGate). No agent decides whether a learning is promoted or a write is allowed.

The four-gate promotion test

Process the techniques and failure modes from completed engagements. Promote a learning to the skill base only if all four hold:

1. Generalizable. Reusable pattern, not target-specific lore. No machine names, lab IDs, target IPs, preserved flags, writeup attributions.
2. Material improvement. Adds coverage, efficiency, or decision-quality for future engagements.
3. Not already captured elsewhere in the skill base. (scripts/skill_linter.py flags duplicates.)
4. Minimal footprint. Prefer extending an existing entry over adding a new file. Keep the base lean and high-signal.

Reframing recipe

Always frame as a reusable pattern: "when encountering X condition, try Y approach" — never "on box-N, Y worked". Use <TARGET_IP>, <DC_FQDN>, <DOMAIN> placeholders in tool examples.

Output

Three buckets, built in code so a run cannot claim an edit it did not make: Updated. / Skipped. (with the gate that failed) / No changes.

Reference

• STRUCTURE.md — directory layout requirements.
• FRONTMATTER.md — YAML rules.
• CONTENT.md — writing guidelines.
• ROUTING.md — technique type → target file.

Anti-Patterns

• Creating CHANGELOG.md / SUMMARY.md / VERIFICATION.md auxiliary files.
• Meta-documentation about the creation process inside the skill itself.
• Verbose inline templates and examples (link to reference/ instead).
• Re-introducing duplicate rule prose (brute-force, output-dir, env-reader).
• Files past their cap — split into reference/ immediately.