AgentSkillsCN

skill_stealer

从任意URL（GitHub仓库、博客文章、推文，或任何描述AI工作流或提示词的来源）中提取可复用的代理技能。按照Agent Skills标准，将核心理念提炼为简洁明了的SKILL.md文件。此外，您还可以“偷取”代码——将脚本重写为Python，并通过uv运行Click CLI。当用户分享一个仓库、技能、提示词或工作流的链接，而您希望将其精髓转化为可复用的技能时，此方法将大显身手。但请勿将其用于一般的网络浏览。

SKILL.md
--- frontmatter
name: skill_stealer
description: >
  Extract a reusable agent skill from a URL (GitHub repo, blog post, tweet,
  or any source describing an AI workflow or prompt). Distills the core idea
  into a clean SKILL.md following the Agent Skills standard. Can also steal
  code — rewriting scripts to Python with Click CLIs run via uv. Use when the
  user shares a link to a repo, skill, prompt, or workflow and you want to
  capture its essence as a reusable skill. Do NOT use for general web browsing.

Given $ARGUMENTS (a URL or description of a skill idea), extract and create a new skill.

Process

  1. •

    Fetch the source - Read the URL content. For GitHub repos: README, prompt files, system prompts, SKILL.md, and any scripts. For tweets/posts: extract the core idea.

  2. •

    Identify the essence - Strip away:

    • •Framework-specific boilerplate (the model knows frameworks)
    • •Verbose explanations of things the model already knows
    • •Marketing language and scaffolding
    • •Over-complicated orchestration that Claude Code handles natively (e.g., bash loops that spawn CLI instances — use the Task tool instead)
  3. •

    Extract the core - What does this tell the AI to do differently? Reduce to the minimum effective prompt. Only keep what Claude doesn't already know.

  4. •

    Choose the right degree of freedom for each part of the skill:

    High freedom (text instructions) — use when multiple approaches are valid, decisions depend on context, or heuristics guide the approach. Most skill content should be high freedom.

    code
    Example: "Validate with the project's quality checks. Fix failures before proceeding."

    Medium freedom (pseudocode, templates, suggested formats) — use when a preferred pattern exists but some variation is acceptable.

    code
    Example: A suggested JSON schema with "adapt as needed for your project"

    Low freedom (exact scripts, specific commands) — use ONLY when operations are fragile, error-prone, or consistency is critical. Rewrite scripts to Python with Click CLIs run via uv run.

    code
    Example: A validation script that must run exactly as written to avoid data corruption

    The test: imagine Claude as a robot on a path. Is this a narrow bridge with cliffs (low freedom needed) or an open field (high freedom, many paths lead to success)?

  5. •

    Steal the code (only when low freedom is justified) - If the source has scripts for fragile/mechanical operations:

    • •Rewrite to Python using Click for CLIs, run via uv run python scripts/<name>.py
    • •Handle errors explicitly (don't punt to Claude)
    • •No magic constants — justify values
    • •If Claude can do it natively (manage JSON, use git, call APIs), don't write a script for it
  6. •

    Format as SKILL.md with YAML frontmatter:

yaml
---
name: extracted-skill-name
description: >
  WHEN to use: <specific triggers>.
  WHEN NOT to use: <boundaries>.
---
  1. •Create the skill directory - Write to skills/<skill-name>/SKILL.md. Scripts (if any) go in skills/<skill-name>/scripts/.

Quality Checks

Apply the Agent Skills Best Practices throughout:

Conciseness: The context window is a public good. Claude is already very smart — only add what it doesn't know. Challenge every line: "Does Claude need this?"

Degrees of freedom: Default to high freedom. Only drop to medium/low when the task is fragile. Over-specifying heuristic tasks wastes tokens and constrains Claude from finding better approaches.

Progressive disclosure: SKILL.md is the overview (under 500 lines). Split reference material into separate files Claude reads on-demand. References one level deep only.

Description drives discovery: Write in third person with specific trigger terms. Claude uses descriptions to choose from 100+ skills. Include both what it does and when to use it.

Workflows for complex tasks: Break multi-step operations into clear sequential steps. Include feedback loops (validate → fix → repeat) for quality-critical operations.

Provide defaults, not options: Don't present multiple approaches unless necessary. Pick the best default, mention alternatives only as escape hatches.

Checklist:

  • •Description uses WHEN/WHEN NOT pattern
  • •Instructions are actionable, not philosophical
  • •Under 500 lines for SKILL.md (split into reference files if needed)
  • •Name is snake_case, memorable, and descriptive
  • •No time-sensitive information

Post-Creation: Compress if Needed

After creating a skill, evaluate its length. If it exceeds 150 lines, run it through the compression process in the skill_pruner skill (Skill Compression section). The key: preserve the procedural skeleton, output schema, key distinctions, and taxonomies. Cut examples, anti-pattern catalogs, genre-specific patterns, meta-commentary, and references. Target ~20% of original length.

Output

Report what you extracted, what you simplified, what freedom level you chose and why, and the path to the new skill files.