antigravity-docs

name	antigravity-docs
description	Look up authoritative, up-to-date Google Antigravity documentation covering the Antigravity 2.0 IDE, the Antigravity CLI, the SDK, agents/subagents, models, skills, rules/workflows, hooks, MCP, browser use, settings, enterprise, migration, and FAQ. Use when the user asks how an Antigravity feature works, what a config field does, how to set up the IDE/CLI/SDK/skills/MCP, when troubleshooting an Antigravity error or unexpected behavior, or when you need to cite current official docs rather than rely on training-cutoff knowledge.
argument-hint	["topic or doc title"]
allowed-tools	Read, WebFetch

Antigravity Docs

Answers questions about Google Antigravity from the official docs, on demand. It reads a local manifest of every doc page, picks the most relevant one(s), and fetches the page's raw Markdown. Always prefer this skill over recalling docs from memory — Antigravity is new and changes fast.

If the user passed an argument ($ARGUMENTS), treat it as the topic to look up. Otherwise infer the topic from the conversation.

When to use this skill

Use it whenever the question is about Antigravity itself or anything in its docs scope:

• Antigravity 2.0 (the agentic IDE): overview, getting started, models, agents/subagents, agent settings, artifact review, skills, rules/workflows, plugins, hooks, MCP
• Antigravity CLI: install, getting started, using, subagents, plugins, reference, migration
• Antigravity SDK, IDE/editor features (tab, side panel, review), browser use, settings, enterprise, migration, FAQ

Why this skill works differently from claude-code-docs / codex-docs

Codex and Claude Code publish .md twins at their page URLs, so those skills fetch a doc URL directly. Antigravity's doc pages are a client-rendered SPA that returns an empty shell — you cannot fetch a page URL and get content. Instead:

• The index is a local file, references/docs-manifest.json, generated by the antigravity-docs-index-builder skill. You Read it — you do not fetch llms.txt or scrape the site.
• The content lives at a separate .md URL (content_url in each manifest entry, under /assets/docs/...). That IS fetchable and is what you WebFetch.

So the flow is: read the local manifest → pick pages → WebFetch their content_url.

Procedure

1. Read the manifest

Read ${CLAUDE_PLUGIN_ROOT}/skills/antigravity-docs/references/docs-manifest.json

It contains _meta (when it was generated, source bundle) and pages, each: {section, slug, title, description, content_url}. If the file is missing, tell the user to run /antigravity-docs-index-builder first — do not try to reconstruct it or guess URLs.

2. Pick the right page(s)

Match the user's question against each entry's title + description, not the slug alone. Then:

• Pick 1–3 pages per batch, not more. The manifest is for triage, not bulk loading.
• One specific feature ("how do Antigravity subagents work?") → one page.
• Cross-concept question ("how do skills relate to rules/workflows?") → pick each relevant page.
• Nothing in the manifest obviously matches → say so. The manifest is the complete list of doc pages; if it's not there, either it doesn't exist or the manifest is stale (suggest a rebuild). Do not invent a content_url.

3. Fetch the batch

For each chosen page, WebFetch its content_url (the /assets/docs/....md URL from the manifest):

WebFetch url=<content_url from manifest>
        prompt="<a question that captures what the user actually needs, not 'summarize this page'>"

4. Evaluate, then loop or answer

After each batch, judge whether the fetched pages actually answer the user's question:

• Enough → answer, grounded in the fetched content. Cite the doc page (title + its antigravity.google/docs/<slug> page URL) so the user can verify.
• Not enough (the answer lives on a page you haven't read, or a page cross-references another) → go back to step 2, pick the next 1–3 pages, and fetch again.
• Keep looping until you can answer, up to a default cap of 9 pages total across all batches.
• Still not enough at 9 pages → stop. Tell the user honestly what you've read, what's still missing, and ask whether they want you to keep reading more pages. Don't silently blow past the cap or pad the answer with guesses.

Rules

• Read the manifest; don't scrape. The list of pages and their content URLs comes only from references/docs-manifest.json. Never fetch a doc page URL (SPA shell — empty) and never guess an /assets/docs/... URL.
• Loop in small batches, cap at 9 pages. Fetch 1–3, check if that's enough, fetch more only if it isn't. At 9 pages still short, ask before continuing.
• A 404 on a content_url means the manifest is stale. The page was renamed or removed upstream. Tell the user to run /antigravity-docs-index-builder to refresh, rather than guessing the new URL.
• Cite the human page, not the asset URL. When citing, link https://antigravity.google/docs/<slug> (what a person can open), even though you fetched the /assets/docs/....md twin.
• Pass through what the docs say. Don't merge aggressively with prior knowledge — the user wants current authoritative behavior, not a synthesis.