Automated Documentation: Using Goose AI Agent to Ship 55 Pages in 4 Days

How I Documented an Entire Product in 4 Days with an AI Agent

Debbie O’Brien used Goose, an open-source AI agent by Block, to document the Zephyr Cloud AI Platform. The project produced 55 documentation pages and 59 screenshots in a single four-day sprint.

Why This Matters

Manual documentation is often the bottleneck in rapid software delivery because maintaining screenshots and consistent tone across dozens of pages is labor-intensive. By utilizing an AI agent with specialized ‘skills’ and a declarative screenshot manifest, developers can treat documentation as a reproducible artifact, allowing for the regeneration of entire asset suites in minutes when UI changes occur.

Key Insights

• Declarative Screenshot Manifests: Defining 59 screenshots in a YAML file allows for automated regeneration when UI changes occur, such as the Zephyr sidebar redesign in 2026.
• Skill-Based AI Specialization: Goose uses markdown-based ‘skills’ to encode style guides, ensuring the agent maintains a consistent voice and formatting across 55 distinct pages.
• OS-Level Automation: Peekaboo facilitates desktop app interaction for Tauri-based apps where standard browser-based tools like Playwright cannot access native webviews.
• LLM-Ready Documentation: Implementing llms.txt and llms-full.txt via Rspress allows AI agents to ingest an entire 3,000-line documentation site in a single request.
• Automated Verification: Playwright CLI is used to verify DOM snapshots of the built documentation to ensure H1 tags and images render correctly before deployment.

Working Examples

A YAML manifest defining navigation steps and validation criteria for automated screenshot capture.

screenshots:
- id: getting-started/app-overview
  output: docs/public/images/getting-started/app-overview.png
  crop: window
  description: >
    Full app window showing the icon rail, channel list,
    and a chat conversation.
  validate:
    - Channels
- id: getting-started/create-channel-dialog
  output: docs/public/images/getting-started/create-channel-dialog.png
  crop: main
  steps:
    - click: '+'
      near: 'Channels'
    - wait: 1.5
  cleanup:
    - press: 'Escape'

Practical Applications

• Use Case: Zephyr Cloud uses the withZephyr() Rspress plugin to upload built sites to an edge network in under 2 seconds for instant peer review. Pitfall: Agents may fabricate URLs with incorrect build hashes unless strictly instructed to grep build logs for real URLs.
• Use Case: Vision framework OCR identifies pixel-accurate bounding boxes for UI text to automate navigation in native macOS applications. Pitfall: OCR failures like misreading ‘update’ as ‘undate’ require substring searches or coordinate-based fallbacks to prevent pipeline breaks.

References:

• https://dev.to/debs_obrien/how-i-documented-an-entire-product-in-4-days-with-an-ai-agent-3338
• https://docs.theaiplatform.app