Lib Skill Writer

Library Skill Writer

Context & Purpose

Use this skill when generating or maintaining AI agent documentation for large, complex, multi-module libraries. The agent does not need a mirror of the codebase; it needs a curated map of how the library behaves at runtime, how to bootstrap and migrate it, which real-world tasks it solves, and how to recover when things go wrong.

This skill is used both for the first-time generation of a skill and for incremental updates when the library changes.

Target Directory Architecture

Every generated skill must follow this file tree and be created inside the current project:

./skills/{skill-name}/
├── SKILL.md
└── references/
    ├── architecture.md
    ├── initialization.md
    ├── migration.md
    ├── recipes/
    │   ├── {use-case-1}.md
    │   ├── {use-case-2}.md
    │   └── ...
    └── troubleshooting.md

Always place the generated skill at ./skills/{skill-name}/ relative to the project root. Do not write skill files to temporary directories, system paths, or locations outside the current project.

1. Root File (SKILL.md)

SKILL.md serves as a router. It contains no implementation details, no class lists, and no method signatures. Keep it short enough to scan instantly.

Routing Map

• Runtime behavior, lifecycle, state, resources, concurrency: references/architecture.md
• Bootstrap and configuration: references/initialization.md
• Version upgrades and deprecations: references/migration.md
• Concrete use cases and tasks: references/recipes/{use-case}.md
• Errors and debugging: references/troubleshooting.md

How to Choose a Reference

Match the user's request to the file name. If no exact match exists, read the closest reference and architecture.md before acting. Never guess based on a module name alone.

Reading Order

1. architecture.md — understand the runtime model and constraints before touching code.
2. initialization.md — if the task involves setup or configuration.
3. recipes/{use-case}.md — the specific task the user wants to perform.
4. migration.md — when migrating code or handling deprecations.
5. troubleshooting.md — only when diagnosing failures.

Critical Constraints

• Do not copy class or method signatures into the skill. Reference the source code or library docs directly when exact API is needed.
• Always verify the current module context before applying a recipe.
• Prefer linking to existing docs over duplicating them.

2. References

References are chunked by user intent. Each file answers one question: what does the agent need to know to do the right thing in this situation?

architecture.md

Capture how the library works at runtime and what invariants the agent must respect.

• Execution lifecycle: initialization, steady state, teardown.
• State machine and resource ownership.
• Concurrency and threading model.
• Resource management rules: connection lifetimes, stream closures, memory leaks, cleanup.
• Module boundaries that matter for reasoning.
• Design principles and invariants the library relies on.

Do not list every class, method, or constant. Focus on concepts that are not obvious from reading the public API.

initialization.md

Capture how to get the library ready for use.

• Required configuration steps and their defaults.
• Environment-specific setup: production, test, development.
• Boot order and dependencies between components.
• Common bootstrap mistakes and how to avoid them.

Keep this file focused on bootstrap and configuration only. Do not use it for arbitrary workflows.

migration.md

Capture how to move between versions or replace deprecated parts.

• Breaking changes between major versions.
• Deprecation paths and replacement patterns.
• Recommended upgrade sequence and rollback options.

Keep this file focused on version changes only. Do not use it for advanced usage patterns.

recipes/

Each recipe is a single, concrete library-specific use case. One file per use case.

File naming rules

• Use kebab-case.
• Start with a verb that describes the user's goal.
• Good examples: configure-connection-pool.md, publish-message.md, handle-retries.md, setup-test-environment.md.
• Bad examples: integration.md, advanced.md, core.md, utils.md.

Recipe boundaries

• A recipe covers one complete outcome for the user.
• Do not split by difficulty. Basic and advanced variants of the same task belong in one file.
• Do not split by technical layer. A recipe follows the user's task end-to-end, even if it touches initialization, runtime, and cleanup.
• If two use cases share steps, cross-reference them. Do not merge unrelated use cases into one file to avoid duplication.

Recipe content

• Why: one sentence explaining why this pattern is used.
• When to use it: context and preconditions.
• Step-by-step instructions.
• Common mistakes and how to avoid them.
• Cross-references to architecture.md for runtime constraints.

troubleshooting.md

Capture how to diagnose and fix failures.

Use a strict table format:

| Error / Symptom | Root Cause | Fix | |---|---|---| | ... | ... | ... |

Each entry must be actionable without reading the source code.

3. Reference Templates

Use these templates to keep generated references consistent.

architecture.md template

# Architecture

## Lifecycle
1. Initialization
2. Steady state
3. Teardown

## Resource Management
- Connections: ...
- Threads: ...
- Memory: ...

## Concurrency Model
...

## Design Invariants
- ...

initialization.md template

# Initialization

## Required configuration
...

## Environment-specific setup
...

## Boot order
...

## Common mistakes
- ...

migration.md template

# Migration

## From v1 to v2
...

## Deprecated features
...

## Rollback options
...

recipes/{use-case}.md template

# [Verb the Noun]

Why: [one sentence explaining why this pattern is used].

When to use it: [context and preconditions].

Steps:
1. ...
2. ...
3. ...

Common mistakes:
- ...

See also:
- Runtime constraints: `architecture.md#[section]`
- Bootstrap or configuration: `initialization.md#[section]`
- Version changes: `migration.md#[section]`

troubleshooting.md template

# Troubleshooting

| Error / Symptom | Root Cause | Fix |
|---|---|---|
| ... | ... | ... |

4. Generation Workflow

1. Create the target directory ./skills/{skill-name}/ inside the current project. All generated files must live under this path.
2. Analyze the library: identify the runtime model, lifecycle, configuration surface, and real-world tasks the library solves.
3. Discover use cases: look at docs, examples, issue trackers, FAQs, and common integration patterns.
4. Cluster by user goal, not by module or technical layer. Name each use case as verb-noun.
5. Write architecture.md first.
6. Write initialization.md and migration.md next — they are required.
7. Write remaining recipe files — one per use case.
8. Write troubleshooting.md from known errors and failure modes.
9. Write SKILL.md last as a router.
10. Verify against the quality gates below.

5. Update Workflow

Use this workflow when the library has changed and the existing skill may need to be updated. Do not regenerate the entire skill from scratch unless explicitly asked.

Step 1: Read the existing skill

Read the current SKILL.md and all files in references/ to understand what is already covered.

Step 2: Analyze the library change

Identify what actually changed in the library: new feature, behavior change, bug fix, performance improvement, deprecation, breaking change, new error mode.

Step 3: Decide if the skill needs an update

Use the decision matrix below.

| Type of change | Does the skill need an update? | Which file to update | |---|---|---| | Internal refactor with no public behavior change | No | None | | New internal class or method, no usage impact | No | None | | New configuration option or bootstrap requirement | Yes | initialization.md | | New runtime constraint, lifecycle step, or resource rule | Yes | architecture.md | | New public API that enables a common use case | Yes | Add recipes/{verb-noun}.md | | New public API extending an existing use case | Yes | Update existing recipes/{use-case}.md | | Existing use case changed or replaced | Yes | Update existing recipes/{use-case}.md | | Use case no longer relevant | Yes | Remove recipes/{use-case}.md and document the removal in migration.md | | Breaking change or deprecation | Yes | migration.md | | New error, warning, or failure mode | Yes | troubleshooting.md | | New routing domain or reading order needed | Yes | SKILL.md |

Step 4: Apply minimal updates

Update only the files affected by the change. Preserve existing structure and wording where possible.

Step 5: Verify

Run the quality gates. Ensure no unrelated files were changed.

6. Quality Gates

Before finishing, confirm:

• [ ] The skill is located at ./skills/{skill-name}/ inside the current project.
• [ ] SKILL.md has no implementation details.
• [ ] initialization.md and migration.md exist at the references/ level.
• [ ] No class or method signatures are duplicated from the codebase.
• [ ] Every file in recipes/ represents exactly one user use case.
• [ ] Recipe file names follow the verb-noun convention.
• [ ] There are no catch-all files like integration.md, advanced.md, utils.md, or recipes.md.
• [ ] initialization.md contains only bootstrap and configuration.
• [ ] migration.md contains only version upgrades and deprecations.
• [ ] Every recipe explains why the pattern is used.
• [ ] Troubleshooting entries are actionable without reading source code.
• [ ] Cross-references between references are explicit.
• [ ] The skill would still be useful if the library gains 30% more API surface.

For updates, additionally confirm:

• [ ] Only affected files were changed.
• [ ] Unrelated references were not rewritten.
• [ ] Removed or obsolete recipes are handled explicitly.

7. Anti-Patterns

Avoid:

• Copying source code or API docs into the skill.
• Putting initialization.md or migration.md inside recipes/.
• Creating catch-all files such as integration.md, advanced-patterns.md, or utils.md.
• Using initialization.md or migration.md as dumping grounds for unrelated workflows.
• Dumping multiple unrelated use cases into one recipe file.
• Naming recipe files after modules, classes, or layers instead of user actions.
• Adding examples/, tests/, or fixtures/ folders — recipes already cover this.
• Treating the skill as a replacement for the library documentation.
• Regenerating the entire skill for every library change.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: bibendi
• Source: bibendi/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.