Content To Skill

Content To Skill

Content To Skill turns messy source material into a runnable skill package. It extracts repeatable mechanisms from content, rejects unsupported claims, and produces a SKILL.md plus any references, examples, scripts, and validation prompts needed for another agent to execute the method later.

Default to Chinese unless the user asks for another language. Prefer creating or patching files when a writable target path is available.

Resource Guide

• Read references/source-intake.md before processing documents, transcripts, videos, multiple files, or unclear source boundaries.
• Read references/skill-construction-protocol.md before writing a SKILL.md, scaffold, references, examples, or validation plan.
• Read references/evaluation-rubric.md when reviewing a generated skill, deciding whether it is publishable, or repairing a weak conversion.
• Run scripts/check_skill_package.py /path/to/generated-skill after creating or materially editing a local skill package.
• Read examples/retest-prompts.md when you need realistic forward tests for a newly generated skill.

First Principle

A source is not the skill. A skill is the compressed control system that lets a future agent repeat the useful behavior without rereading the whole source.

Every conversion must preserve this chain:

source evidence -> reusable mechanism -> executable workflow -> output contract -> validation signal

Core Workflow

1. Route The Request

模式选择: choose the smallest mode that preserves the source evidence and produces an executable artifact.

Choose the smallest mode that fits:

If multiple modes match, prefer the one with stronger evidence requirements: REPAIR_FROM_FEEDBACK, PACKAGE_BUILD, SOURCE_AUDIT, QUICK_CONVERT.

1.5. CHECKPOINT / STOP Gates

Emit CHECKPOINT / STOP and wait before:

• overwriting an existing skill folder with unrelated local edits
• publishing or pushing a generated skill to a public repository
• processing private, paid, sensitive, or inaccessible source material
• expanding one source into a router, skill family, or SkillBank
• claiming full_test quality without actually running an independent or fresh-context test

The checkpoint must name the target path or repository, files affected, validation command, rollback point, and exact approval needed.

2. Establish Evidence Boundary

Identify what was actually read and what remains unavailable.

Capture:

• source type: text, document, PDF, spreadsheet, slide deck, audio, video, repository, web page, notes, or mixed
• accessible artifacts: pasted text, local paths, transcripts, extracted text, screenshots, metadata, or summaries
• unavailable artifacts: video without transcript, unreadable binary files, private links, missing attachments, or unsupported formats
• user goal: new skill, skill family, prompt workflow, repair, or evaluation
• target runtime: Codex skill folder, single SKILL.md, repository scaffold, or patch plan

For video or audio, first obtain a transcript, chapter notes, subtitles, or user-provided summary using available tools. If none is available, ask for the transcript or permission to extract it. Never infer detailed methods from a title alone.

3. Extract Mechanism Cards

Do not summarize the source paragraph by paragraph. Extract portable controls:

## Mechanism Card
Name:
Source evidence:
Trigger:
User job:
Decision rule:
Procedure:
Output:
Quality signal:
Failure mode:
Skill location:
Keep / merge / discard:

Keep only mechanisms that change future agent behavior. Discard anecdotes, slogans, duplicate examples, and unsupported claims unless they become examples, boundaries, or tests.

4. Choose Skill Shape

Select one production shape before writing files:

• Single procedural skill: one bounded repeated task.
• Engineering workflow pack: coding, debugging, architecture, tests, issues, or PR workflows.
• Methodology toolbox: router plus focused diagnostic skills.
• Cognitive distillation skill: a person, worldview, creator, company, or field turned into a perspective lens.
• Skill family or SkillBank: many independent skills with routing, acquisition, versioning, or state.

Default to a single procedural skill unless the source clearly requires routing, shared state, multiple domains, or independent subskills.

5. Generate The Skill Package

When a writable path is available and the user asks to create or convert, create files before the final answer.

Minimum package:

skill-name/
├── SKILL.md
├── agents/
│   └── openai.yaml
└── examples/
    └── retest-prompts.md

Add references/ when the source has detailed protocols, rubrics, evidence notes, examples, or domain facts that should not bloat SKILL.md.

Add scripts/ only when repeated mechanical checks, extraction, conversion, or validation should be deterministic.

For public GitHub repositories, add only the human-facing files needed for reuse:

README.md
LICENSE
.gitignore
test-prompts.json

Keep the skill package itself portable: SKILL.md, agents/, references/, examples/, and scripts/ must still work if copied into any skills-compatible runtime.

Generated SKILL.md must include:

• frontmatter with only name and description
• overview and first principle
• resource guide with direct links to references and scripts
• routing or operating modes
• stable workflow that produces decisions or artifacts
• output protocols
• boundaries and anti-patterns
• quality standard and validation loop