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