# Gemini > Use Gemini ACP for grounded search/research, citations, source discovery, supplied-document recall or gemini prompt/extract/summarize/translate/code-review. Use when this capability is needed. - Skill: `tomevault-io/gemini-26` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add tomevault-io/gemini-26` - Raw SKILL.md: https://api.skillmd.com/api/skills/tomevault-io/gemini-26/raw - Safety review: pending (external: skill-scanner PASS, skillspector PASS) - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Product & Planning - Author: tomevault-io (https://skillmd.com/u/tomevault-io) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/tomevault-io/gemini-26 --- # Gemini Role: ACP-powered Gemini skill for grounded search/research, citations, source discovery, supplied-doc recall, and Gemini prompt/extract/summarize/translate/code-review. Scope: Discovery + synthesis via `pi-gemini-acp`. Does NOT invoke scraper tools directly — call scraper separately when needed. MUST NOT bypass auth/CAPTCHA/blocks, claim hidden extension-to-extension execution, or imply directory scans/credential-file access. MUST reject unsafe paths in `gemini_analyze`; requires filesystem-read/resource-link capability. Base64 images are validation-only. Local/no-key mode works only over supplied docs/sources. ## Tools - `gemini_status` — ACP command/auth/capability preflight. - `gemini_ask` — prompt/extract/summarize/translate/code-review supplied text; no path reads or edits. - `gemini_search` — web grounding or supplied local-doc search. For latest/current/news/time-sensitive queries set `bypassCache: true`; recall reuse is opt-in with `useRecall: true`. - `gemini_research` — source/citation research with optional safe fetch. Use `bypassCache: true` for latest/current/news topics; `useRecall: true` only when reuse is acceptable. - `gemini_analyze` — explicit local file/image paths only; validates paths, rejects unsafe paths, requires filesystem-read/resource-link capability. Base64 images are validation-only. - `gemini_results` — get stored output by `responseId` as an overview/source/raw paged view, or run local SQLite FTS recall. - `/gemini-config cache|recall|status|command|permissions|trust` — inspect/configure ACP, cache, recall, permissions, and Gemini CLI trust. - Optional scraper: `web_scrape`/`web_batch` for reading URLs found by Gemini; `web_map`/`web_crawl` only for site-structure tasks. ## Workflow - `gemini_status` if ACP readiness matters. - `gemini_search` for URL discovery; `gemini_research` for source/citation synthesis. - For fresh/current/latest/news requests MUST pass `bypassCache: true`; MUST NOT opt into recall unless user asks to reuse prior work. - Prefer primary/high-authority sources. If scraper tools exist and exact claims matter, scrape top URLs before answering. - Cite URLs. Distinguish Gemini snippets/citations from scraper-verified page text. - `gemini_analyze` for user-specified files/images only; never imply directory scans or hidden/credential-file access. - Use `gemini_results(action: "get", responseId)` for stored-result overviews; use `view: "source"` with `sourceId` or `view: "raw"` only when more detail is needed. - `gemini_results(action: "recall")` as honest local FTS recall; zero hits are normal. ## Scrape after Gemini when - exact quotes, dates, numbers, or claims matter; - snippets are thin/conflicting; - the URL is likely canonical docs, paper, changelog, release note, policy, or source page; - the user asks to verify/audit/compare/extract. Skip scraping for quick link lists, one supplied-text summary (`gemini_ask`), inaccessible/private pages, or when snippets are enough. --- > Source: [brandonkramer/pi-gemini-acp](https://github.com/brandonkramer/pi-gemini-acp) — distributed by [TomeVault](https://tomevault.io).