study

这段内容是审核时保存的快照。外部来源才是完整且最新的版本。

---
name: study
description: "A personal, cross-project learning queue and tutor. Capture topics you want to understand deeply into a dumb home-dir backlog (~/.study/topics.md) from anywhere, then on demand run a guided, canvas-driven, recall-checked deep dive on any one of them. Topics can be grounded in your real code, in a sandbox the agent scaffolds, or be purely conceptual (\"investing\", \"solving a rubik's cube\"). Use whenever the user wants to learn or understand something deeply, park a rabbit hole for later, or be taught/quizzed — \"add X to my study list\", \"I want to understand Y\", \"teach me Z\", \"quiz me on…\", \"let's do a deep dive\", \"what's on my study queue\", \"explain how our <thing> actually works\" — and as the place other skills (e.g. better-planning-sync, better-planning-comprehend) send knowledge gaps. Works standalone; uses the canvas skill for interactive lessons when installed."
---

# Study

Relying on agents to write code (and to answer everything) quietly erodes understanding: the
rabbit holes you'd once have chased — "wait, how *does* RRULE expansion work?" — get skipped to
stay on task, and the gap compounds until you're a spectator to your own systems. `study` is the
counter-habit. It splits learning into two cheap halves: **capture now** (a frictionless backlog of
things worth understanding) and **learn later** (a guided, tested deep dive when you actually have
the time). It's general — code is its strength, not its limit; "investing" and "the Krebs cycle"
belong on the same queue.

## The home: a dumb queue, the filesystem as state

Everything lives under `~/.study/` (full protocol in `references/study-layout.md` — read it before
touching the files):

- **`~/.study/topics.md` is pure backlog** — one topic per line, nothing else. A bare topic, or a
  topic plus free-form context if the author felt like typing it. **No status tags, no dates, no
  machine fields, no parsing.** It stays short and human; anything (including you, by hand) can
  append a line.
- **A topic's state *is* its directory.** No `~/.study/<slug>/` ⇒ not started (just a queue line).
  Directory exists ⇒ picked up. Done ⇒ recorded *inside* the dir (the learning record marks it).
  Read state by looking at the filesystem, never by parsing the queue.
- **Picking a topic up graduates its line out** of `topics.md` into a fresh `~/.study/<slug>/`
  workspace that seeds the dive. The queue shrinks as you act; a topic is never in two places.

## The four verbs

The user's intent maps to one of these — do the one they asked for; don't force a dive when they
only wanted to capture.

### capture — append a line
"Add X to my study list", or a topic surfaced mid-conversation worth parking. Append one line to
`~/.study/topics.md` (create the file on first use). Keep the human's words; attach free-form
context only if it's cheap and useful (e.g. the repo + path the question came from), never machine
tags. Confirm in a sentence; do **not** start teaching. This is the GTD capture — get it out of the
head and move on.

### browse — show the queue, help pick
"What's on my study queue?" Show `topics.md` (and, if asked, the started/finished dirs from
`ls ~/.study/*/`). Help choose what to learn now — by what's quick, what's blocking real work, or
what the user's in the mood for. Recommend one, but it's their call.

### learn — scaffold and run the dive
The user picks a topic (from the queue or fresh). Derive a slug, scaffold `~/.study/<slug>/`, move
the topic line + its context out of `topics.md` into the workspace seed, then run the dive (next
section). If a workspace already exists for the slug, resume from its learning record instead of
starting over.

### record — capture what stuck, mark done
On the way out of a dive (or when the user says they're done), write/append the learning record:
what clicked, what's still fuzzy, the next thing to chase. Marking it done lives **in the dir**, not
the queue. The record is both the retention aid