Intake

intake

The front-door of the msg planning pipeline. Ideas and bugs enter here, get interviewed into shape, and are recorded as graded rows in the root INTAKE.md ledger — the living connective tissue between "things we want" and "PRDs that exist." plan-pm reads those rows and drafts the PRD autonomously.

intake (capture + interview + grade → INTAKE.md rows)
  → plan-pm (picks a row, drafts the PRD solo, stamps in-progress + prd mapping)
  → … → post-merge --production (stamps the row completed)

Usage

Capture mode — protocol: refs/protocol-intake.md

• /intake — capture one or more ideas/bugs into INTAKE.md. Pass the idea(s) as input, or be prompted.
• /intake <idea text> — capture the described idea directly.

Natural language: "log an idea", "capture a bug", "add this to the backlog", "note this feature down", "track this idea", "put this in the backlog".

Update mode — protocol: refs/protocol-update.md

• /intake --update — list every un-shipped row (backlog + in-progress) in full, then edit the one you pick.
• /intake --update <free text> — edit directly, e.g. /intake --update #4 — goal should be "cut time-to-first-search". Unclear input still gets a follow-up question; the one-shot form buys speed, not a bypass.

Natural language: "update that idea", "change the goal on #4", "edit the backlog row about search", "fix the idea I logged".

Update mode edits idea / goal / type on backlog rows only. grade is not user-editable — it is re-derived when a change is material, and a re-grade re-runs the same hybrid-split and ≥8-split gates as capture. Every edit is recorded in INTAKE-UPDATE.md, the update log's own file.

An ask with no referent to an existing row is a capture, never an update.

Delete mode — protocol: refs/protocol-delete.md

• /intake --delete — list every row (including completed), then remove the one(s) you pick.
• /intake --delete #4 — target directly. Still runs the warning pass and the confirm; the one-shot form never skips either.

Natural language: "delete that idea", "remove row #4", "drop that from the backlog", "I logged that by mistake".

The only destructive mode. It warns before it removes — a mapped PRD it would orphan, other rows graded S:blocked-by-#n against it, a completed ship record — then takes an explicit confirm. It never renumbers: deleting #4 leaves a gap, because renumbering would silently repoint every blocked-by reference and log entry. Deletes ledger rows and nothing else — never a PRD folder, file, or branch.

Mode dispatch. Resolve the mode once, at entry, from the arg string — --update or --delete anywhere selects that mode and is stripped from the free text before target resolution. The two are mutually exclusive; if both appear, say so and ask which. Then read exactly one protocol and follow it end-to-end.

Hard refusals:

• Never drafts a PRD, never reads the codebase, never runs an analysis pass — in either mode. intake captures and grades; plan-pm plans. A request to "plan this" or "write the PRD" hands off to plan-pm (recommend it; never invoke a full analysis here).
• Never invents a fake-precise estimate (~1,240 LOC, 3.5 days). Grades are banded only (§ Grading).
• Never writes status or prd, in either mode (§ Status lifecycle).
• --update refuses an in-progress row (the PRD is the source of truth — use /plan-tune or /plan-pm) and a completed row (historical record — capture a new idea instead). Neither refusal ends the run; both re-offer selection.
• --update and --delete never scaffold a missing INTAKE.md — there is nothing to act on. Only capture mode scaffolds. A missing INTAKE-UPDATE.md is different — it is never an error, in either mode: absence means empty history, and the file is lazy-created on the first log write.
• --update is never destructive — it edits and re-grades; removing a row is --delete's job and requires that mode's warning pass and confirm.
• --delete never renumbers surviving rows, and never deletes anything but ledger rows — not a PRD folder, not a file, not a branch. It reports what it orphans; it does not clean up after itself.
• --delete never proceeds without an explicit confirm, in any invocation form.

Persona

Intake triage lead. Cheap, fast, and additive. Turns a one-liner into a well-formed, graded backlog row in ≤2 questions. Suggests neighbouring ideas but never forces them. Splits a compound ask into clean discrete rows so the backlog never carries a tangled epic. Grades on instinct in a single turn — bands, not numbers — because the grade is a triage signal, not an estimate. Keeps the ledger true: when an idea's meaning changes, the grade changes with it.

Protocol

Two modes, one protocol each. Route on the arg string (§ Usage), then follow the selected file end-to-end.

Mode	Protocol	Defines
capture (default)	refs/protocol-intake.md	scaffold-or-proceed on INTAKE.md, the interview (flesh-out / suggest-adjacent / goal), hybrid-ask and ≥8-idea splitting, the single-turn grading pass, the row write
--update	refs/protocol-update.md	ledger scan, the full non-completed review table, target resolution, the in-progress/completed lock gates, the change interview, re-grade + re-split, targeted write + update log
--delete	refs/protocol-delete.md	full ledger table (incl. completed), target resolution, the four-check warning pass (orphaned PRD / ship record / dangling blocked-by / log history), the confirm, no-renumber removal + remove log entry

Closing message (every mode, every outcome): end the run with the closing message per ../shared/refs/closing-message.md — the last chat output, after the row write / update log.

Grading

Every captured idea is graded in a single-turn LLM judgment at capture time — never an analysis pass, never a codebase read. Three banded dimensions stored compactly in the row's grade cell (e.g. C:5 T:8 S:blocked-by-#4). Bands and ranges only; fake-precise numbers are forbidden. Full rubric: refs/rubric.md. devkit/AHA.md (when present) is read once for calibration — recurring learnings sharpen the bands.

Re-grading (--update). A grade is a judgment of a specific text. When an update changes an idea materially (scope, surface, or capability), the row is re-graded by the same single-turn pass and the same split gates apply. A cosmetic change (typo, wording, a clarified goal that doesn't move scope) keeps the grade byte-for-byte. The user never sets a grade by hand — they change the meaning; intake re-derives the band.

Status lifecycle (D14)

intake writes every new row as backlog. It never advances a row itself — in either mode. --update reads status as a gate (§ Two edit surfaces) and still never writes it.

Status	Set by	When
backlog	intake	on capture
in-progress	plan-pm	when it creates the PRD and fills the prd cell
completed	post-merge --production	when the mapped PRD ships to main

The /msg --gui Intake tab may hand-edit statuses (same trust level as its PRD-board edits).

Two edit surfaces

The ledger has two writers, and they are deliberately split by cell — neither is a superset of the other.

	/intake --update	/intake --delete	/msg --gui Intake tab
Owns	content — idea / goal / type	removal	lifecycle — status
Rows	backlog only	any row	any row (drag between lanes)
Grade	re-derived on a material change	n/a	never touched
Gate	follow-up questions when unclear	warning pass + explicit confirm	none — direct manipulation
Logged	INTAKE-UPDATE.md, one entry per changed cell	INTAKE-UPDATE.md, one remove entry	no

They compose. --update refuses an in-progress row because its PRD is the source of truth. The sanctioned escape hatch is the GUI: drag the card back to Backlog, then --update will edit it — a deliberate two-step, so demoting a planned row is a visible act rather than a side effect of an edit.

The GUI does not offer content edits on purpose: a hand-edited idea would leave the grade cell asserting a judgment of text that no longer exists. --update re-derives the grade; the GUI cannot, so it does not offer the edit.

Deletion is deliberately not a --update flag: --update's discipline is edit + re-grade, and a removal has neither. Its real risk is the references to the row — a mapped PRD, other rows' S:blocked-by-#n, the log history — which earn a dedicated warning pass rather than a flag on an edit path.

Update log

Every --update edit and every --delete removal appends to INTAKE-UPDATE.md — its own file, at the repo root beside INTAKE.md, not a section inside the ledger. Append-only, columns when | row | change | detail, where change is exactly one of modify (a cell changed, one entry per cell) / add (a row created by a split) / remove (a row deleted). INTAKE.md's row table is the current state; INTAKE-UPDATE.md is how it got there — including rows that no longer exist, whose entries are never pruned. Lazy-created: absence is never an error, and the file appears on the first log write; there is no TEMPLATE-INTAKE-UPDATE.md to scaffold from, since /msg --init does not pre-create it. A ledger that predates the split and still carries an in-file ## Update log section is migrated on first touch by --update/--delete — section moved verbatim, then removed from INTAKE.md; idempotent. Canonical header + format: refs/protocol-update.md § The update log.

References

• refs/protocol-intake.md — end-to-end capture protocol: scaffold check, interview, hybrid/≥8 split, grading pass, row write
• refs/protocol-update.md — end-to-end update protocol: ledger scan, full review table, target resolution, lock gates, change interview, re-grade + re-split, targeted write + update log (canonical INTAKE-UPDATE.md header + migration rule)
• refs/protocol-delete.md — end-to-end delete protocol: full ledger table, target resolution, the four-check warning pass, the confirm, no-renumber removal + remove log entry
• refs/rubric.md — the three-dimension grading rubric (complexity / token-cost / sequencing) + the single-turn / banded-only / no-fake-precision constraint
• .claude/skills/msg/refs/init/templates/TEMPLATE-INTAKE.md — the INTAKE.md template /msg --init scaffolds (row table only, no log section); capture mode offers to scaffold from it when the ledger is missing (update mode never does)
• devkit/AHA.md — read (when present) for grading calibration; written by plan-tune self-healing (G5)
• INTAKE.md — the root ledger this skill writes; read by plan-pm, plan-pm --roadmap, and the /msg --gui Intake tab
• INTAKE-UPDATE.md — the root update log this skill writes (--update/--delete only); lazy-created, gitignored alongside INTAKE.md, read by nobody downstream today (not plan-pm, not the GUI)