| name | browser-fallback |
| description | Use a deterministic browser workflow only when direct fetch or extraction is insufficient for a public-web task. |
| origin | repo-local web research workflow |
Browser Fallback
Use this skill when static fetches fail because a page is JS-heavy, stateful, lazy-loaded, or otherwise not inspectable enough without a real browser session.
Workflow
- Confirm that repo search, direct fetch, and
site-extract are insufficient first.
- Use the harness-provided browser tool or the most deterministic browser stack available.
- Reuse one session per site, inspect DOM and network behavior before adding extra actions, and record the exact final URL.
- Capture the minimum reproducible navigation steps, selectors, and visible result state.
- Once the data is visible, hand the content back to
site-extract and source-audit.
Stop Rules
- Do not bypass CAPTCHAs, paywalls, rate limits, bot checks, geo blocks, or access controls.
- Do not brute-force forms, rotate identities, or simulate evasive traffic.
- If login or a human challenge is required, stop and ask for a manual session, a permitted export, or an alternate source.
- Avoid destructive actions, account changes, or purchases.
Good Defaults
- Prefer deterministic Playwright/CDP-style actions over vague browsing.
- Reuse the same browser tab/session when iterating on one site.
- Keep notes precise enough that another maintainer can replay the path.
Read next
docs/reference/ai-agent-web-research-stack.mdsite-extract, source-audit, and compliance-and-rate-limit
