Skip to content
GitHub

Getting Started with Product Documentation

Document your own repository with the author-product-docs skill: audit it, read the journey gap report, write the first missing page, and check it.

In about 20 minutes you will find out which pages your project’s docs are missing, write the most important one, and check it against your code.

With agentbundle installed, add the pack:

agentbundle install --pack product-documentation

Open your agent in the repository you want to document and send this request:

Audit this project's docs and tell me which stages of the reader journey are missing

You should see the agent start reading your repository. The audit changes no files.

  • A repository with something users run or call: a library, CLI, HTTP API, app, service, framework, plugin, or agent-context pack.
  • An agent that supports installed packs, such as Claude Code.
  • agentbundle, the installer. See install routes.

The author-product-docs skill starts by working out what your product is. It looks for evidence: a package manifest and public API for a library, command parsers for a CLI, an OpenAPI or protobuf file for an API, a configuration schema for a service. A repository can ship more than one surface, and the skill documents each one a reader touches.

It then looks for where your repository keeps user-facing docs and maintainer docs, using the agent-guidance file, the contributing guide, and the existing layout. It reads your README and your docs index — a docs site’s navigation config counts as the index — and samples a large doc set by stage. It runs no project code unless you say you trust the repository; this tutorial assumes your own repository, so tell it so if you want it to run checks.

You should see the report name the surface it found. If the evidence fits none of the known surfaces, the skill says what it found and asks once.

The reply starts with a table of nine rows, one per stage a reader passes through:

StageReader’s question
Discover and evaluateIs this for me?
InstallCan I get it running?
First successCan I make it do one real thing quickly?
Daily tasksHow do I do this specific thing?
Look upWhat exactly does this accept or return?
UnderstandWhy does it work this way?
TroubleshootIt broke. What do I check?
UpgradeWhat changed, and what must I change?
ContributeHow do I report or change something?

Each row is covered, partial, missing, or not applicable, with the file that shows it or the reason. A library with no outside contributors may mark Contribute not applicable. Nothing marks First success not applicable.

The next actions are ranked by where readers are lost first. A missing or failing first success comes before a missing explanation. The top-ranked row marked missing or partial is the page to write first.

Ask the skill to write the page for the top-ranked row. For example, when First success ranks first:

Write a quickstart for this project

For example prompts for other stages, see How to author product docs.

The skill reads the canonical sources for your surface before it writes. It then drafts one page by default. You should see the page open with the goal and a first runnable action in its first 120 words, and the draft should say what the product reads and what it may change, show the expected result, and end with a link to the next stage. The skill writes the page where your repository already keeps docs of that kind, and asks once if it cannot tell where that is.

You decide whether the page matches what you want readers to do first.

Ask the skill to verify what it wrote:

Verify this page against what the code does

You should get three lists: verified claims, unverified claims, and claims that contradict current behavior. Then run the check that fits your surface. Run the examples for a library. For a CLI, compare the page with --help and run each command. For an API, compare it with the contract file. For a service, compare it with the configuration schema. For an app, walk each documented task in the running app or its end-to-end tests. For a plugin, compare the page with the contribution block in its manifest. For a framework, compare each extension-point entry with its interface and the code that loads it. For an agent-context pack, read each skill’s source and send the first starter prompt.

Fix any contradiction before you publish. A claim nothing checked stays labeled unverified.

Send the audit request from the top of this page again. The row you fixed should now read covered, with the new page as evidence. Work down the list one page at a time.

  • A journey gap report for your project’s docs.
  • One new page, checked against the code, that links to the stage after it.
  • A repeatable loop: audit, write the top gap, verify.