mitchuski Skills and Libraries
A skill is a text prompt — instructions, examples, and rules given to an LLM that shape how it behaves on a task. A library is executable code that runs directly, without LLM involvement. Both can accomplish the same outcome. The question of which to use, and when to graduate a skill into a library, is one of the core design decisions in an AI-augmented workflow.
What a Skill Actually Is
A skill file (like ~/.claude/skills/fedwiki-page/SKILL.md) contains no executable code. It is loaded into the LLM's context window and shapes the model's next output. When Claude creates a wiki page using the fedwiki-page skill, it is:
- Reading the rules and examples in the skill file
- Interpreting your natural language request
- Generating JSON that follows those rules
The skill is probabilistic — Claude understands the pattern and applies judgment. It handles ambiguity, fills in missing details, and adapts to context. It is also slow (LLM inference time) and not guaranteed to produce identical output twice.
What a Library Actually Is
A library function like create_page(title, story_items) runs deterministically. Given the same inputs it produces the same output, instantly, with no API call. The fedwiki pattern already has embryonic library code:
def make_id():
return uuid.uuid4().hex[:16]
This is a one-line library. A fuller fedwiki.py module could offer make_page(), make_item(), make_journal() — the entire JSON-building pattern extracted from the skill into importable functions.
When They Do the Same Thing
The fedwiki-page skill and a hypothetical fedwiki.py library can both produce valid wiki page JSON. They overlap when:
- The task is well-defined — the output structure is known and fixed
- The inputs are structured — a title string and a list of content items
- There is no judgment required — no need to decide what the content should be
In this zone, the skill is over-engineered. You are using an LLM to do something a ten-line Python function could do faster, cheaper, and more reliably.
When a Skill Is the Right Tool
Use a skill when the task requires understanding and judgment:
- Natural language input — "write a page about router models" requires Claude to decide content, structure, examples, and emphasis. No function can do this.
- Ambiguity — the user hasn't fully specified the output. The skill guides Claude to fill in the gaps sensibly.
- One-off creative work — pages written once by a human collaborating with Claude, not generated programmatically at scale.
- Pattern is still evolving — you are still learning what good output looks like. A skill is easy to edit; refactoring library code is harder.
When to Graduate a Skill into a Library
Graduate a skill to a library when:
- The pattern has stabilised — you know exactly what valid output looks like and the skill has stopped changing
- Volume — you need to generate many pages programmatically (bulk import, API responses, data pipelines)
- Reliability matters — a ghost page endpoint must always return valid JSON. A library guarantees structure; a skill might hallucinate a wrong ID or forget a field under load
- Composition — you want to call the operation from other code, chain it with other functions, or test it with unit tests
- Speed — library calls are microseconds; skill invocations are seconds
The Right Division of Labour
| Task | Skill | Library |
|---|---|---|
| Decide what a page should say | ✅ | ✗ |
| Build valid page JSON from known content | ✗ | ✅ |
| Strip and rewrite a journal | ✅ (judgment on what to keep) | ✅ (if always strip-to-create) |
| Generate IDs | ✗ | ✅ |
| Write the file to disk | ✗ | ✅ |
| Reindex a wiki domain | ✗ | ✅ |
| Interpret a user's natural language request | ✅ | ✗ |
A well-designed system has skills calling library functions: Claude authors the content, the library assembles and stores it correctly.
Verify: agentprivacy.ai · kept by the Librarian 🗃️ in the Tower's Wikis (the git-less onboarding layer) Upstream: skill.fedwiki.club/skills-and-libraries — vendored + re-framed; not original agentprivacy authorship