notebook

name	notebook
description	Use when the user wants to capture, review, or close the open loops Claude surfaced during the current conversation — the insights, suggested next steps, and follow-up questions Claude raised while chasing something else, which get buried as the conversation branches. Persists them to `.smart/notebook.md` as a running list of unfollowed leads. Trigger on `/smart:notebook`, or when the user says "notebook", "记录本", "做题本", "open loops", "what did we not follow up on", "别丢", "被掩埋", "未跟进", "现在有哪些没跟进的". A companion `Stop` hook already auto-captures marked blocks (★ Insight, 建议下一步) after every reply; this skill adds the free-form leads the hook cannot parse and manages their open/closed status. This is NOT `todo` (either/or decisions under a pinned mainline) and NOT `distill` (reusable knowledge Q/A archive) — notebook tracks unclosed leads so a question chased into another question never gets lost.
argument-hint	empty = mine the conversation & list open leads · done <id> = close one
model	sonnet

notebook — Claude's open-loop notebook (做题本)

What this is — and how it differs from todo / distill

As you chase one question, Claude surfaces others — an ★ Insight, a "建议的下一步", a follow-up question. The conversation branches, and those leads get buried. The notebook is a running list of open loops: leads Claude raised but that have not been followed up on yet.

It is deliberately a third thing, orthogonal to its siblings:

• todo tracks either/or decisions under one pinned mainline.
• distill archives reusable knowledge Q/A into a topic knowledge base.
• notebook (this) tracks unclosed leads — open loops — so nothing gets lost as the conversation wanders.

Content language follows the conversation.

The two-layer design (why a hook AND a skill)

Capturing open loops must not depend on Claude remembering to record them — the moment the conversation branches is exactly when Claude is most likely to forget. So capture is split across two layers:

• A Stop hook (hooks/notebook-capture.py) already runs after every reply, deterministically — it cannot be skipped and does not rely on Claude's self-discipline. It scrapes the two block shapes Claude emits in a fixed format — the ★ Insight … ───── border block and a "建议的下一步 / Suggested next steps" heading section — and appends them to .smart/notebook.md.
• This skill is the intelligent half. The hook is a dumb scraper: it catches only the marked blocks and cannot judge a free-form suggestion or a mid-sentence follow-up question. When invoked, this skill re-reads the conversation, adds the leads the hook missed, and manages their open/closed status.

Together: the hook guarantees nothing marked is ever lost; the skill guarantees completeness and lets you close loops.

State file: .smart/notebook.md

One fixed file per project, git-ignored (personal scratch, never committed). The hook creates it if absent. Shape:

# 做题本 · Notebook
<!-- 自动由 Stop hook 捕获 · /smart:notebook 管理 · updated <YYYY-MM-DD HH:MM> -->

## 🔵 Open（CC 抛出、尚未跟进的线索）
- [ ] N1 <lead title>  · 🕒 <ts> · 来源:insight <!-- h:xxxxxxxxxx -->
  > <excerpt / gist>
- [ ] N2 <follow-up question>  · 🕒 <ts> · 来源:notebook

## ✅ Closed（已跟进 / 已回答 / 已采纳）
- [x] N0 <title> → 已跟进 (<date>) · <one-line outcome>

Each Open entry carries a stable N# id and a hidden <!-- h:hash --> fingerprint the hook uses for dedup. Leads this skill adds itself use 来源:notebook (no hash needed — the skill dedups by meaning).

Argument

$0	Action
(empty)	Sync + list — mine the current conversation for open loops, reconcile into .smart/notebook.md, print the Open list.
done <id>	Close — move one lead (e.g. done N2) from Open to Closed with a one-line outcome.

Empty: sync + list

1. Resolve the file. Project root via git rev-parse --show-toplevel (fallback: cwd); the notebook is <root>/.smart/notebook.md. Read it if present — the hook may already have populated it.
2. Mine the current conversation — the user + assistant messages of THIS session. Collect open loops: insights, suggested next steps, and follow-up questions Claude raised but that were never actually pursued or answered. Deliberately include the free-form ones the hook cannot parse — a suggestion phrased in prose, a question asked mid-paragraph. Do not re-mine source-file TODO comments or the native task list — those are different things.
   • A lead is "open" only if it was not followed up. If Claude raised it and the conversation already resolved it, it belongs in Closed, not Open.
3. Reconcile each lead against existing entries — by meaning, not wording (this discipline is borrowed from todo):
   • New — nothing matches → add under Open with the next N# id and 来源:notebook.
   • Duplicate — an existing entry (including a hook-captured one) already covers it → leave it untouched; count it as already tracked. Match against the hook's entries too, so you never re-add what it already caught.
   • Diff — matches an existing entry but adds detail → append the new bit to that entry; do not spawn a second one.
4. Write back, then refresh the updated stamp.
5. Print the Open list to the conversation, most-worth-following-up first — a compact view so the user can see what is still dangling. Never write the report itself into any file other than the notebook.

done <id>: close a loop

Move the entry with that id from Open to Closed: flip [ ]→[x], append → 已跟进 (<date>) · <one-line outcome>. Never delete it — closing is a state change, so a later-reopened question keeps its history.

Guardrails

• Never delete a lead. Only append, or flip open→closed. When unsure whether a lead is closed, keep it Open — a dangling entry is cheap; a lost lead is the whole thing the user is trying to avoid.
• Share the file with the hook safely. Both the hook and this skill write .smart/notebook.md — always read-modify-write, and preserve entries you did not touch, including the hook's <!-- h:... --> markers (strip one and the hook will re-add that lead as a duplicate).
• Do not overwrite the hook's work. Treat hook-captured entries as first-class; dedup against them rather than rewriting them.
• The report is printed to the conversation only; the file holds the leads.
• This skill owns exactly one file (.smart/notebook.md) and no .smart/settings.json keys.