IDEA-CAPTURE — a fleeting idea that is not captured is lost

An idea that arrives mid-work and is not written down is gone by the next session. This skill is the store for raw ideas: one Markdown file per idea, in the project's ideas/ directory, written verbatim and never edited. Calibration is a separate step (idea-calibrate); this skill only makes sure the idea is still there when you are ready for it.

Rules

• Capture first, calibrate later. When a raw idea arrives, write it down before doing anything else with it — before answering, before coding, before the next message.
• Verbatim, always. The raw text is the idea as spoken, with the date and who said it. Never edit, improve, or interpret it after capture.
• The store lives in the project (ideas/), committed with it. A store in /tmp is lost on the next restart; a store in the project survives.
• A dropped idea is never deleted. Dropping only changes the status and records the reason. The file, the raw text, and the reason stay.
• Statuses move one way: raw → calibrated → scoped, or any → dropped. A dropped idea never comes back as raw; if it returns, capture it again as a new idea.

Steps, in order

1. Capture

Run the script (scripts/capture.py, Python 3.11+, standard library only):

capture.py capture --store <project>/ideas --text "<the idea, verbatim>" --source "<who said it>"

It prints the idea id (YYYYMMDD-HHMMSS-slug). The slug is ASCII only (a-z, 0-9, hyphen): the raw text may be any language — a Persian-only idea gets the slug idea (with a -2 suffix if two collide in the same second) — while the raw text itself is stored verbatim. An empty idea is refused (exit 1, no file created). If two ideas arrive in the same second, the second gets a -2 suffix — the id is always unique. The raw text is stored exactly as given; only the emptiness check strips whitespace, and CRLF line endings are normalised to LF. The frontmatter carries raw_length (the length of the raw text), so an idea that itself contains a line reading ## History is not truncated. The source must be a single line (a newline would inject a frontmatter field). The id is strictly validated (only the generated shape is accepted) and the resolved path must stay inside the store, so an id can never escape it — even through a symlink. Listing and searching resolve every file and require it to stay inside the store, and a file whose frontmatter id does not match its filename is reported as malformed. A status change takes a lock file (created with O_EXCL; the holder's PID is written into it, so a lock left by a dead process is recognised as stale and removed with a note, while a lock with no readable PID is never stolen and a lock is never removed for age) for the whole read-modify-write, then writes atomically (temp file + replace), so two concurrent changes cannot lose a History entry and an interruption cannot leave a half-written file. A capture creates its file exclusively, so two concurrent captures cannot overwrite each other. A file that cannot be parsed is reported to stderr, never silently skipped — and a list or search that skipped one exits 2 (partial), so a machine caller never treats an incomplete result as complete. A status change edits only the frontmatter's status: line, never a status: line inside the raw text, and a file with no status recorded fails cleanly instead of assuming one.

2. Find it later

• capture.py list --store <project>/ideas — every idea, with status and date.
• capture.py list --store <project>/ideas --status raw — only the unprocessed ones.
• capture.py search --store <project>/ideas <keyword> — ideas whose raw text contains the keyword (case-insensitive; works in Persian and English). Search matches only the raw idea text, not the frontmatter or the History, so searching for "raw" or "dropped" does not return everything.
• capture.py show --store <project>/ideas <id> — the whole file.

3. Move it forward

When the idea is calibrated with idea-calibrate, record it:

capture.py promote --store <project>/ideas <id> calibrated --note "<what it became>"

When the calibrated brief is scoped into v1, promote it to scoped. When the idea is not going ahead, drop it — with a reason, always:

capture.py drop --store <project>/ideas <id> --reason "<why not>"

Every move appends one line to the file's History. The raw text is never touched.

Report

Return the report in this form:

• Captured (id, date, source).
• Store (path, number of ideas by status).
• Moved (every status change, with the reason).
• Verdict: PASS (the idea is on disk verbatim, findable by search, and every move is recorded) or FAIL (an idea was lost, edited, or deleted, or a move was not recorded).

Scope

This skill is the store only. It does not sharpen, validate, or scope an idea — that is idea-calibrate. It does not decide what to build — that is the user, at the Idea → Code gate. The store is plain Markdown files; there is no database, no server, and no sync — the project's own git history is the backup.

Sources

Written from scratch. No ready skill answers this need: the research found capture-in-session (idea-calibrate, step 1) but no persistent store for fleeting ideas between sessions.