Skip to content

SKILL.md

Latest commit

 

History

History
49 lines (37 loc) · 3.13 KB

SKILL.md

File metadata and controls

49 lines (37 loc) · 3.13 KB
Raw
name ask
description Ask the second brain in plain words — semantic search over the curated knowledge vault (embeddings + reranker), not a full-corpus dump. Trigger on "/ask <question>", "what do I have on <topic>", "what did I write about X". Token-cheap: RAG retrieves the smallest relevant slice, the LLM only synthesizes the top hits.
license MIT

/ask — query the Second Brain

🧒 When reporting to a non-technical operator: end with a child-simple "In plain words" recap in their language.

Deterministic-first (the operator's token law): RAG retrieves the smallest relevant slice; synthesize ONLY that — never dump the corpus.

Wrong lane? /ask = semantic MEANING over the curated vault. For exact WORDS inside chats → /search; an exact PERSON name → /find; a CHAT by name → /chat (all 0 tokens, deterministic). Use those when the question isn't conceptual.

🖥️ Visual search (a live server — the operator works visually)

For interactive eyes-on search: python "$IMPORTS_ROOT/ask_server.py" (or start_ask.bat) → open http://127.0.0.1:8770. It loads e5 + reranker ONCE, after which each query takes ~3s. A search box + filter chips (my writing only / concepts / insights / leads / people / conversations / notes), cards with the rerank score, type, date and a ⏳ staleness flag. CPU by default (does not fight the GPU fleet, holds 0 VRAM); --gpu when the GPU is free. This is the GUI for the operator; --ask below is for me synthesizing an answer in chat.

Run (CLI — for synthesizing an answer in chat)

python "$IMPORTS_ROOT/brain_ask.py" "<question>" Scope with filters (cheaper + sharper):

  • --anton (only the owner's own writing) · --concepts (distilled "what I think about X") · --insights
  • --person <name> · --conv (conversations) · --leads (CRM) Returns top-K chunks (chunked + tagged + edit-aware index); may flag ⚠ STALE on volatile facts >90d.

Answer

  • Synthesize the returned hits into a DIRECT answer; cite the note titles so the operator can open them.
  • If a hit is ⚠ STALE, say so and offer to re-verify (per the epistemic-decay layer).
  • If retrieval is thin/irrelevant → say so, suggest a sharper query or a filter; do NOT pad with guesses.
  • Keep it tight; end with 🧒 recap.

Note

The index is kept fresh by the reindex routine. If results feel stale right after a big import, mention a reindex may be due (brain_embed_update.py) — but don't reindex unprompted.

About & contact

Built and battle-tested at Palo Alto AI Research Lab — a fleet of Claude Code machines running 24/7 as a second brain and synthetic cofounder. Every skill here survived real production use before publication.