Imported from pradeepmouli/skillit (packages/typedoc-plugin/skills/skillit-docs/SKILL.md). Install upstream with npx skills add pradeepmouli/skillit --skill skillit-docs. Copyright stays with the author.

Documentation Conventions for Skill Generation

typedoc-plugin-skillit generates AI agent skills (SKILL.md) from your TypeScript documentation. The quality of the generated skill depends entirely on what you write in your source docs.

The generator runs an audit during pnpm typedoc that checks your docs against these conventions at four severity levels:

• fatal — skill is broken without this (fix first)
• error — LLMs can't write correct code from this
• warning — skill works but isn't as good as it could be
• alert — stylistic suggestion

Fixing Audit Issues

Run pnpm typedoc and look for the audit output after [skills] lines. Fix fatals first, then errors.

Fatal — Fix These First

F1: package.json description — One-sentence problem statement:

{ "description": "Generate AI agent skills from TypeScript API documentation" }

NEVER: taglines ("Fast and lightweight"), tech-stack descriptions ("A TypeScript library").

F2: package.json keywords — 5+ domain-specific words:

{ "keywords": ["proxy", "ipc", "child-process", "worker", "type-safe", "rpc"] }

NEVER: generic terms (typescript, library, npm, utils) — these are filtered out and add zero triggering value.

F3: README description — Blockquote after title:

# my-lib

> Map process.env to strongly-typed nested config objects with camelCase fields.

F4: JSDoc on every export — At least one sentence:

/** Render a single extracted skill into SKILL.md + progressive disclosure references */
export function renderSkill(skill: ExtractedSkill): RenderedSkill {

Error — Fix These Next

E1: @param descriptions — Describe what each parameter controls, not what type it is:

// GOOD:
/** @param options Configuration for rendering — controls token budgets and output format */

// BAD:
/** @param options The options */

E2: @returns on non-void functions:

/** @returns Rendered skill with SKILL.md discovery file and on-demand reference files */

E3: Interface property JSDoc:

export interface Config {
  /** Maximum token budget per reference file (default: 4000) */
  maxTokens: number;
}

E4: At least one @example — Realistic usage, not trivial:

/**
 * @example
 * ```typescript
 * import { renderSkills, writeSkills } from '@skillit/core';
 * const rendered = renderSkills(extracted, { maxTokens: 4000 });
 * writeSkills(rendered, { outDir: 'skills' });
 * ```
 */

E5: package.json repository — Add the repo URL.

Warning — Polish for Quality

W1: Add @packageDocumentation to index.ts with overview and feature list. W2: Add @example to every exported function. W3: Use @deprecated, @since, @throws, @see tags where applicable. W4: Add 10+ domain-specific keywords. W5: Add ## Features section to README. W6: Add ## Troubleshooting section to README — common issues, errors, and FAQs users encounter.

JSDoc Tags for Expert-Quality Skills

These three custom tags are what separate a "passing" skill from a genuinely useful one. They provide the expert knowledge that Claude doesn't already have — decision procedures, anti-patterns, and freedom calibration.

The plugin auto-registers these as blockTags at load time — no typedoc.json config needed. If they're accidentally placed in modifierTags, the plugin auto-moves them with a warning (modifierTags strips content).

@useWhen — When to reach for this class/function

Each bullet is a scenario where an agent should use this export. Only write entries that contain non-obvious expert knowledge — if Claude could guess the answer from the class name, don't write it.

BAD (obvious): @useWhen - Displaying images on Sprite — Claude already knows Sprite displays images. GOOD (expert): @useWhen - Rendering >500 identical sprites — ParticleContainer batches into a single GPU draw call vs O(n) individual draws

For large libraries (50+ classes), write @useWhen on 5-10 key decision points, not every class. Focus on scenarios where the wrong choice causes performance cliffs or silent bugs.

/**
 * @useWhen
 * - You need CPU-intensive work isolated from the main process — IPC overhead is ~1ms per call
 * - You want EventEmitter events forwarded transparently across process boundaries
 */

@avoidWhen — When NOT to use this, and what to use instead

Use the — delimiter to name the alternative:

/**
 * @avoidWhen
 * - The class uses non-serializable state (closures, WeakMaps) — use worker_threads instead
 * - You need sub-millisecond latency — IPC adds ~1ms overhead, use in-process calls
 */

@never — Anti-patterns from experience

NEVER + BECAUSE + FIX format. Always include the non-obvious reason AND a recovery path.

/**
 * @never
 * - NEVER pass functions as constructor args — V8 serialization silently drops them. Fix: use sanitizeV8: true or restructure to pass data only
 * - NEVER call $terminate() inside a proxied method — creates IPC deadlock. Fix: return from the method first, then terminate from the caller
 */

@throws — Error recovery paths

Standard TSDoc. Include the recovery path, not just the error type. This is where "what to do when things go wrong" lives.

/**
 * @throws {ModuleResolutionError} When the class module path cannot be detected from static imports. Fix: pass an explicit modulePath as the second argument
 * @throws {SerializationError} When constructor args contain functions or Symbols. Fix: use sanitizeV8: true or restructure to pass plain data only
 * @throws {TimeoutError} When INIT handshake exceeds timeout. Fix: increase timeout option or check that the child module exports the class
 */

@remarks — Decision trees, trade-offs, thinking frameworks

Standard TSDoc. Use for the expert reasoning that helps an agent CHOOSE between options. On @packageDocumentation, this becomes a thinking framework in the SKILL.md body.

/**
 * @remarks
 * ## Choosing a serialization mode
 * - `json`: lower overhead, simpler debugging, works in all test runners
 * - `advanced`: supports Buffer, Map, Set, BigInt, TypedArray
 * - Use `json` unless you specifically need non-JSON types
 * - Use `advanced` + `supportHandles: true` only for OS handle passing (sockets, pipes)
 *
 * ## Token budgets
 * Applied per-reference-file, not per-skill. A skill with many functions
 * may have functions.md truncated but types.md will be unaffected.
 */

@category — Group exports by domain

Groups exports in SKILL.md Quick Reference and reference files. Overrides the automatic filename-based grouping.

/** @category Rendering */
export function renderSkill(...) {}

/** @category Rendering */
export function writeSkills(...) {}

/** @category Token Management */
export function estimateTokens(...) {}

Complete Example

/**
 * Run class instances in isolated child processes.
 *
 * @remarks
 * All method calls become async via IPC. V8 serialization is 10x faster
 * than JSON but doesn't support functions or Symbols.
 *
 * @useWhen
 * - You need CPU-intensive work in a separate process
 * - You want EventEmitter forwarding across processes
 *
 * @avoidWhen
 * - The class has non-serializable state
 * - You need sub-millisecond latency
 *
 * @never
 * - NEVER pass functions as constructor args — silently dropped by V8
 * - NEVER call $terminate() inside a proxied method — IPC deadlock
 *
 * @param target Class constructor to proxy
 * @param modulePath Path to module exporting the class
 * @returns Proxied instance with all methods returning Promises
 * @throws {ProcxyError} When child process fails to start
 * @see sanitizeForV8 — check if values survive serialization
 * @category Core
 */

Impact on Generated Skill Quality

The remaining gap to Grade B requires hand-tuned prose quality — but the tags create the scaffolding.

README Structure

The generator extracts specific sections by exact heading name:

# package-name

> One-sentence problem statement.

Extended explanation with key capabilities.

## Quick Start

` ` `typescript
import { main } from 'pkg';
main();
` ` `

## Features

- **Feature** — what it does and why

## Troubleshooting

- **Error X occurs when** — [cause and fix]

Accepted heading variants (case-insensitive):

Why This Matters (Skill-Judge Alignment)

Generated skills are evaluated on 8 dimensions. The biggest gaps without conventions:

• Knowledge Delta: Function signatures are things Claude already knows. The value comes from @param prose, @returns meaning, @throws conditions, and @never entries.
• Anti-Patterns: Without a ## Troubleshooting section, skills score ~2/15 on anti-pattern quality.
• Description: Without proper package.json description + keywords, skills never trigger.

Content Priority

The generator pulls content from multiple sources. When the auto-generated skill is thin on a dimension, add content at the highest-priority source:

Every skill-judge dimension maps to a specific tag — add content at the right source:

• D1 Knowledge Delta low? → Add @remarks with trade-offs and decision trees
• D2 Procedures low? → Add @packageDocumentation @remarks with "before X, ask yourself..." frameworks
• D3 Anti-Patterns low? → Add @never with NEVER + BECAUSE + FIX rules
• D8 Usability low? → Add @throws with recovery paths + @example with worked code

Workflow

1. Run pnpm typedoc — see audit output and score estimate
2. Fix fatals first (package.json description, missing JSDoc)
3. Add @example to key exports (surfaces as Quick Start and worked examples)
4. Add @useWhen/@avoidWhen/@never to key exports (biggest quality jump)
5. Add @remarks to @packageDocumentation — architectural context and thinking frameworks
6. Fix errors — @param/@returns on non-self-documenting params
7. Re-run — verify score improvement
8. Iterate until audit estimate plateaus
9. Mandatory: Run skill-judge on the generated skills/<name>/SKILL.md — the audit estimate overestimates by 10-20 points. Skill-judge catches: token bloat, missing thinking frameworks, description trigger quality, orphaned references.

A skill is not ready to ship until skill-judge confirms Grade B (96+) or higher.