| name | browser-automation-playbook |
| description | Operate interactive websites with Prometheus browser tools, including navigation, clicking, forms, uploads, downloads, dynamic-page extraction, screenshots, and browser UI verification. Use for live browser interaction; do not use for ordinary web research, native desktop apps, or X-specific workflows that have their own skill. |
Browser automation
Use browser tools for interactive website state. Use web_search, web_fetch_batch, or web_fetch for normal reading and research; use desktop tools for native apps.
Core loop
- Open or observe the page and use the returned snapshot.
- Choose the narrowest observation that supports the next action.
- Act with a current DOM reference when possible.
- Verify meaningful state changes before continuing.
- Close only sessions Prometheus created when cleanup is appropriate.
Prefer observe:"compact" for orientation, observe:"delta" for incremental UI changes, observe:"snapshot" when the next step needs refs, and screenshots when visual layout or canvas state matters. Do not request another snapshot when the previous action already returned sufficient state.
Input and extraction
- Use fill-style actions for normal inputs and type-style actions for editors, search boxes, and controls that depend on key events.
- Use browser-native upload/download actions rather than manipulating OS dialogs.
- Use structured extraction for repeated records with a known schema.
- Use page-text collection for prose and broad reading inside a dynamic page.
- Use scroll collection only when pagination or ordinary extraction cannot retrieve the data.
- Use JavaScript or network interception as a fallback inspection route, not as the default interaction path.
Recovery
When a ref is stale, the page navigated, or a click appears ineffective, stop and refresh the observation. Do not repeat blind actions. If the DOM is sparse or misleading, capture fresh visual evidence and use vision-guided interaction. If authentication requires an existing user-owned browser profile, confirm the appropriate browser surface instead of silently switching profiles.
For direct assets, use the direct download tool. For supported social/video pages, use the media download path. Verify file existence and content before reporting success.
Read details only when needed
- Read detailed-guide.md for advanced browser tools, observation modes, download/media decisions, and recovery patterns.
- Read session-hygiene-browser-close.md before deciding whether to close a session.
- Read the relevant topic under
references/workflows/ only for the matching workflow; do not load all historical recipes.
Finish with evidence from the current page or downloaded artifact, not an assumption based on the requested action.
