← ClaudeAtlas

grounder-searchlisted

Search this project's vault content for relevant context.
andrej-kolic/grounder · ★ 0 · AI & Automation · score 69
Install: claude install-skill andrej-kolic/grounder
Search this project's vault content for relevant context. **Silence:** write **no assistant text** until step 3. Rounds 1–2 are tool calls with an empty/absent text part — not “I’ll search…”, not “I’ll read…”, not “**Analyzing…**”, not query/terms narration. Use this when the user asks to find prior project-vault context by topic, keyword, concept, or phrase. Scope is this linked project only — the CLI resolves and searches under the linked project vault root. Do not search outside it. ## Turn budget (speed) **Rounds 1–2: tool calls only — no text part in those messages.** Not even one sentence. Exactly **two** assistant turns with tools, then the answer. Allowed tools, nothing else: 1. Round 1 — Shell only: one `search … --json` (optional second search in the *same* round only per the broaden rule below). Message = that tool call, nothing else. 2. Round 2 — Read only: **one** parallel batch of full-file reads (hits 1–4). Message = those Read calls, nothing else. 3. Final answer to the user (first and only chat text). **Do not** Glob, Grep, extra Shell, or status/UI tools (`UpdateCurrentStep`, `TodoWrite`, and similar). Do not add a third tool turn. Do not explore the repo. ## Output contract (default — hybrid) - Do not echo commands or shell output. - **Never paste CLI JSON, snippets, or raw stdout into chat** — parse `--json` internally only. - One final synthesized response only. **Path links (mandatory for every listed file):** - Visible title = `hits[].relativ