agentsclimarketplace

Documentation writer

Skill balacodeio/balakit/skills/documentation-writer

Opinionated, cross-agent rules and skills for AI coding agents — one npx install into Cursor, Claude Code, Codex, OpenCode & Copilot.

Install
npx -y skills add balacodeio/balakit --skill documentation-writer

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Write clear, scannable documentation using a research-first, phase-separated workflow. Covers READMEs, doc sites, API references, migration guides, and technical markdown. Uses web search for up-to-date facts and delegates broad repo reads to an explore subagent before drafting. Use when the user asks for documentation, technical writing, READMEs, doc websites, API docs, or readable markdown — or when accuracy, versions, or external APIs matter.

SKILL.md

6.1 KB, as published. Nobody here has run it

Documentation Writer

Leading words: research-first, evidence-aware, phase separation, scannability, BLUF (Bottom Line Up Front), progressive disclosure.

Research-first technical documentation workflow. Each phase ends with a checkpoint the agent must complete before proceeding — phase separation prevents the agent from drafting prose before the truth is gathered.

How this relates to subagents

Cursor does not expose user-defined nested sub-sub-agents in skill files. What you can do in practice:

  • Phased work: Treat each phase below as a distinct pass. That mimics routing without a separate UI.
  • Built-in delegation: If the host agent can launch a Task / explore subagent, use a read-only explore pass for large or unfamiliar codebases before drafting. Do not assume that tool exists; use it only when present.
  • Online research: Use web search (and official docs links from results) for anything version-sensitive or external.

Phase 1 — Discover (know what exists in-repo)

Read relevant files. For unfamiliar or large codebases, delegate broad mapping to an explore subagent when available.

Exit criteria: You can list every file/API/feature the docs must cover. 🛑 Checkpoint: Do not proceed to Research until the user confirms the scope (or you are explicitly operating solo and the scope is unambiguous).

Phase 2 — Research (external truth)

Run web research before asserting facts about:

  • Library/framework APIs, CLI flags, config keys, breaking changes, deprecations
  • Install commands, system requirements, LTS/support timelines
  • Security or compliance claims (cite sources; avoid unsourced absolutes)

Process: Form 2-4 precise search queries (product + version + topic). Prefer primary sources (official docs, repo README, standards bodies) over random blogs. Capture citations (link + short note of what it supports). If sources conflict, say what is uncertain and what to verify locally.

Exit criteria: Every claim you will make has a source link. 🛑 Checkpoint: Do not proceed to Outline until Research is complete.

Phase 3 — Outline (structure before prose)

Headings, audience, prerequisites, nav/sidebar plan for sites. Apply BLUF: lead each section with the direct answer in 40-60 words, then expand. Use progressive disclosure — summary first, depth below.

Exit criteria: The document skeleton is approved. 🛑 Checkpoint: Present the outline. Do not draft until confirmed.

Phase 4 — Draft (write)

Apply the style rules below. Start each section with the BLUF answer in the first 40-60 words. Use HTML tables for comparisons, definition lists for specs. Lead with concrete examples.

Exit criteria: Every section from the outline has content.

Phase 5 — Verify (correctness)

Re-read against the repo. Re-check any critical external claims. Run the code examples if possible.

Exit criteria: Every code example runs as documented; every external claim still matches the cited source.

Core principles

  1. BLUF — Bottom Line Up Front. Lead each section with the direct answer.
  2. Progressive disclosure — summary first, depth below. Readers (and LLMs) extract the gist immediately.
  3. Scannability — headings, bullets, tables. Every paragraph earns its place.
  4. Evidence-aware — cite sources for versions, APIs, install commands.
  5. Concrete over abstract — examples before explanations.

Standard documentation structure

When creating a new file or README:

  1. Title + 1-2 sentence summary — what it does and why it exists (BLUF).
  2. Quick Start / Installation — minimum steps to get running, exact commands.
  3. Usage / Examples — show, don't tell. Most common use case first.
  4. API / Reference (if applicable) — parameters, options, returns. Tables for structured data.

Formatting guidelines

  • Sentence case headings. Do not skip levels (H1 > H2 > H3).
  • Language-tagged code blocks (```typescript).
  • BLOCKQUOTE for callouts (> Note, > Warning).
  • HTML tables for comparisons, <dl> for specifications.
  • Descriptive link text ("Read the setup guide"), not "Click here".

Style rules

  • Active voice ("The script generates a file").
  • Direct address / imperative for instructions ("Run this command").
  • Consistent terminology throughout the document.
  • One core idea per paragraph. LLMs extract passages, not pages.

Workflows

Creating a documentation website

  1. Recommend a framework if not specified (VitePress, Nextra, Docusaurus, MkDocs).
  2. Establish a clear sidebar/navigation structure: Introduction/Getting Started → Core Concepts → Guides/Tutorials → API Reference.
  3. Write content with each page focused on a single topic.

Reviewing existing documentation

  1. Check for outdated information; web-search claims that depend on versions, APIs, or third-party behavior.
  2. Improve formatting for scannability.
  3. Simplify complex sentences.
  4. Ensure code examples are correct and well-formatted (match the repo or cited docs).

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.