braindump

name	braindump
description	Split long-form text into atomic inbox notes. Accepts inline text or file paths. Triage later via /note process.
allowed-tools	Agent Bash Read

braindump

Split long-form text (planning, retros, design ramblings) into atomic thoughts. Chunks land in notes/ for later triage via /note process.

Vault I/O

Writes inbox notes via the CAPTURE pipeline in _shared/capture-pipeline.md. All vault writes flow through obsidian CLI. Read is used for non-vault input file ingestion (text/markdown paths).

Image routing

If any image paths are present in the argument list → read ${CLAUDE_PLUGIN_ROOT}/_shared/image-capture.md before parsing input.

Vault path

See §1 Vault path resolution. If no vault is configured, abort with No vault configured — run /wiki init first.

Input parsing

Positional argument(s) — inline text and/or file paths:

1. Empty or whitespace-only → abort: /braindump requires text or a file path.
2. Parse input as space-separated text snippets and/or file paths.
3. For each path argument, resolve: absolute when <arg> starts with /; otherwise relative to <vault_root> (not CWD).
4. Path resolves to a readable text file → use file contents as body.
5. Path resolves to a readable file that is neither a supported text file nor a supported image type → abort: Unsupported input type: <ext>. /braindump accepts text, markdown, and image inputs.
6. Path does not resolve:
   • If <arg> looks like a supported image input by extension → abort: Image not found or unreadable: <path>
   • Otherwise → treat <arg> verbatim as inline text. No error.

Split — atomic-thought rubric

Atomic thought = one self-contained idea. Split when topic/claim/referent shifts. Do not split mid-argument or merge distinct claims. Preserve content verbatim; only boundaries are chosen.

Zero chunks (unexpected empty result from the reasoning step) → hard-abort, no retry: /braindump split returned no chunks. Original text not captured.

CAPTURE loop

When to dispatch agents vs. run inline

Chunks	Input order matters?	Mode
1	n/a	Inline — run CAPTURE directly on the main thread.
2–4	yes (sequential argument)	Inline — chunks must run in order so K can MATCH-append to K-1.
2–4	no (independent thoughts)	Agent fan-out — dispatch one agents/capture.md per chunk in parallel.
5+	no	Agent fan-out — always parallel for 5+ chunks.
5+	yes	Inline — sequential order must be preserved; run in order.

Order matters for numbered lists, narratives, build-on-each-other arguments. Independent observations/questions/tasks can run parallel.

Inline CAPTURE (sequential)

For each chunk in order, re-enumerate <vault_root>/notes/*.md fresh (so chunk K can MATCH-append to a note written by chunk K-1). Then:

1. MATCH/NEW per §4 — skip notes/index.md and status: deferred; cap at 20 most recent.
2. MATCH or NEW path per §4; slug via §3.
3. Index patch per §6.
4. Record filename + success/failure. On error: append to failure list, continue — never abort the loop.

Agent fan-out (parallel)

When dispatching agents, verify CWD first:

cd "${VAULT_ROOT}" && pwd   # confirm vault root before agent fan-out

Dispatch one agents/capture.md per independent chunk. Pass each agent:

• CHUNK — the verbatim chunk text
• VAULT_ROOT — $VAULT_ROOT
• SOURCE_PROJECT — basename(cwd)
• TODAY — ISO date YYYY-MM-DD

Wait for all agents to complete. Collect their Filed: / Appended: / Error: lines.

Note: parallel agents cannot MATCH-append to each other's notes (they run concurrently). If two chunks would logically MATCH the same note, run them inline in order instead.

After all agents finish: apply a single consolidated notes/index.md patch. For each Filed: line, prepend one checkbox row under ## Pending:

- [ ] YYYY-MM-DD [<source_project>] <title>

Agents do not patch the index; the orchestrator owns that write.

source_project = basename(cwd). Frontmatter: note shape from §2, no braindump provenance.

Confirmation output

Captured N notes:            ← "note" singular when N=1
- notes/YYYY-MM-DD-<slug>.md
…

No NEW/MATCH labels. No diff. No reasoning.

If any chunks failed:

Captured N notes:
- notes/YYYY-MM-DD-<slug>.md
…

Failed: K chunks.            ← "chunk" singular when K=1
- <one-line reason per failure>
…

Examples

Multi-thought:

user> /braindump I keep forgetting to check the lint score before PRs. Also need to revisit the hot cache size — it's been growing. And the daily skill confirmation message looks wrong in dark mode.
assistant>
Captured 3 notes:
- notes/2026-04-27-check-lint-score-before-prs.md
- notes/2026-04-27-hot-cache-size-growing.md
- notes/2026-04-27-daily-skill-confirmation-dark-mode.md

Single thought (no spurious split):

user> /braindump The slug truncation rule needs to account for multi-byte unicode characters — right now it can split in the middle of a grapheme cluster, which breaks vault filenames on some filesystems.
assistant>
Captured 1 note:
- notes/2026-04-27-slug-truncation-multibyte-unicode.md

File input (vault-relative):

user> /braindump retro-notes/2026-04-feature-retro.txt
# resolved as <vault_root>/retro-notes/2026-04-feature-retro.txt
assistant>
Captured N notes:
- notes/2026-04-27-…
…

One chunk fails:

assistant>
Captured 2 notes:
- notes/2026-04-27-check-lint-score-before-prs.md
- notes/2026-04-27-daily-skill-confirmation-dark-mode.md

Failed: 1 chunk.
- notes/: permission denied writing 2026-04-27-hot-cache-size-growing.md