| name | nestbrain-ask |
| description | Query the user's NestBrain knowledge base from inside Claude Code, using natural language. Use whenever the user says variants of "cerca sulla mia knowledge …" / "search my knowledge …" / "what do I have in my knowledge about …" — with an optional project scope ("progetto X" / "project X"). Returns retrieved articles synthesized into a grounded answer with `[[wikilink]]` citations. |
nestbrain-ask
Run a NestBrain knowledge search and answer from the cited articles. Be terse. No narration, no "let me try …", no fallbacks beyond what's written here.
Steps — execute in order, do not improvise
1. Resolve CLI invocation (one shot, do not announce)
Try in this order, use the first that resolves:
command -v nestbrain succeeds → use nestbrain.- The installed NestBrain.app ships the CLI. Test these paths (first that exists wins) and use it directly via
node:
- macOS:
/Applications/NestBrain.app/Contents/Resources/web/apps/web/nestbrain.bundle.cjs
- Windows:
%LOCALAPPDATA%\Programs\NestBrain\resources\web\apps\web\nestbrain.bundle.cjs
When found, the invocation is node "<that path>".
- Dev source:
<nestbrain_root>/Projects/nestbrain/packages/cli/src/index.ts exists → npx tsx "<that path>".
- None of the above → reply "NestBrain CLI not found." and stop.
Don't say what you tried. Just use what works.
2. Parse user input
- Question: the part after "cerca sulla mia knowledge" / "search my knowledge" / "what do I have in my knowledge about" (and after the optional project clause).
- Project: the value following "progetto X" / "project X" / "per il progetto X" / "from project X", if present.
3. Run ONE search
<CLI> search [-p <project>] "<question>" -l 5
Run from the NestBrain root (cd <nestbrain_root>) so the CLI auto-resolves the workspace paths.
Do not try alternate phrasings, translations, or broader queries. One search. Done.
4. Branch on the result
Zero results → reply (matching the user's language):
Non ho trovato nulla nella knowledge. (IT)
Nothing in the knowledge for this. (EN)
Then stop. Do not grep Library/Knowledge/ directly. Do not read files. Do not speculate. Do not list what you tried.
Any results → read the top 2–3 results' files at <nestbrain_root>/Library/Knowledge/<filePath>, then synthesize a concise answer.
5. Answer (only when results were found)
- Match the user's language.
- Stick to what the read articles actually say. No invention.
- End with a
**Fonti:** (IT) or **Sources:** (EN) line listing the citations as [[<slug>]] (the filename without .md).
Hard rules
- No nested Claude calls. Never use
nestbrain ask — only nestbrain search. The synthesis is your job.
- No "thinking out loud". No phrases like "provo con", "let me also try", "verify the index". One search, then either the answer or "non ho trovato nulla". A user reading the transcript should see the result, not the reasoning steps.
- No direct grep / file walks as a "fallback" when search returns nothing. If the CLI says nothing was found, that's the answer.
- Scope is restrictive. When the user named a project,
-p <name> filters to that project only. When they didn't, omit the flag — search everything.
