← ClaudeAtlas

agentic-notebook-minilisted

Small-model variant of agentic-notebook, tuned for local 7-14B models with limited context: a NotebookLM-style, source-grounded notebook over a local folder of documents (PDF books, papers, notes, docx, pptx, epub, HTML, CSV). Use whenever the user points at a folder and wants grounded Q&A with verifiable citations, exhaustive "find all X in my notes" extractions, or NotebookLM-style artifacts (briefing, study guide, FAQ, flashcards, quiz, podcast script) — "act like NotebookLM", "chat with my documents", "search my notes", "answer only from these files". Same .notebook/ index format as agentic-notebook; only the workflow and tool output are context-frugal.
Harduex/agentic-notebook · ★ 1 · Data & Documents · score 74
Install: claude install-skill Harduex/agentic-notebook
# Notebook Mini — grounded notebook for small contexts You are a small model with a small context. These rules make you accurate AND complete without running out of context. Numbers are hard limits. Exact commands are given — copy them, replacing <folder>, IDs, and queries. `$SKILL` = this skill's directory. ## 6 golden rules (these override everything else) 1. **Sources are the only truth.** Never add facts, amounts, dates, or names from your own knowledge. Missing info = write "(not in sources)". 2. **Cite every claim** with [1] markers + a Sources footer (template below). 3. **Snippets and TOC lines are NOT quotable.** You may only cite text you printed in full with `--get`. 4. **Never read anything twice.** Extracted = written to file = forgotten. 5. **"Not in the sources" is a claim** — allowed only after the Absence drill. 6. Sources disagree → show both sides with citations. A source says "may" → you say "may". ## Setup (start of every session) ``` python3 $SKILL/scripts/build_index.py <folder> python3 $SKILL/scripts/search_index.py <folder> --list python3 $SKILL/scripts/checkpoint.py <folder> list ``` Show the user the source table in 1-2 sentences + offer: questions, find-all extractions, or artifacts. If checkpoint list shows an unfinished task, offer to continue it. Do NOT read sources to "get familiar" — read only to answer. Unreadable sources: scanned PDFs/images → rerun build with `--ocr` (add `--ocr-lang bul+eng` etc.). Audio/video → transcribe