requirement-intake

Turns an untagged task on the board — a paragraph of intent — into an approved, executable plan before any code is written. Covers the workflow/ scratch folder, the questions protocol, sizing, and wha

SKILL.md

GitHub

Requirement Intake

A tagged task ([backend] Add /health endpoint with test) means you already did the requirement engineering — it goes straight to the task-workflow skill.

An untagged task is a requirement. It gets a requirement engineer first.

- [ ] TASK-31 Reply digest for operator — status: todo
      Operator gets a daily email digest of new replies grouped by niche.
      Link each reply to the business detail page.

Title on the line, description indented. That is the only thing a human writes.

One pen

The human writes in tasks.md and nowhere else. They may read anything.

Everything an agent produces goes in workflow/ — a gitignored scratch folder at the repo root. Two surfaces, one owner each, no ambiguity about who wrote what.

<repo root>/
  tasks.md              human writes here. git-tracked. short. the index.
  workflow/             agents write here. gitignored. verbose. the working memory.
    .env                knobs (self-created on first run)
    TASK-31-reply-digest/
      state.md  plan.md  assumptions.md  questions.md
      review-1.md  review-2.md  backlog.md

workflow/ is organised by slice, never by file type

One task = one folder = every artifact for that task, together.

✗ workflow/plans/   workflow/reviews/   workflow/state/     grouping by file type
✗ workflow/backend/ workflow/frontend/                      grouping by agent
✓ workflow/TASK-31-reply-digest/                            grouping by task

The second one matters: a requirement usually spans modules and agents, so it has no home module. The task is the slice.

Bootstrap (first run, idempotent)

  1. workflow/ missing → create it.
  2. /workflow/ missing from .gitignore → append it. Never rewrite the rest of the file.
  3. workflow/.env missing → write it with these defaults and comments:
# task-agent workflow knobs. Per developer, gitignored.
PLAN_GATE=medium          # never | small | medium | always — sizes that need your approval
AUTO_MERGE=off            # off | gated  (gated conditions are in the task-workflow skill)
MERGE_TARGET=develop
MAX_REVIEW_ROUNDS=3       # reviewer FAIL -> fix -> re-review, this many times, then blocked
MAX_FIX_CYCLES=2          # gate failures fed back to the implementer per round
NOTIFY=osascript          # osascript | print | <shell command taking the message as $1>

Unreadable, missing, or unparseable .env → use these defaults with AUTO_MERGE=off. The one irreversible action always fails closed.

The intake phases

Intake is one phase: a run sets status: planning, runs the requirement engineer, and does not stop until the task has landed on its next human gate or gone straight to implementation. planning is a crash marker, never a resting place — a task found sitting in planning had its run interrupted, and the next run restarts the phase.

board status what runs next
todo (untagged) create workflow/<TASK-id>-<slug>/ + state.md, set planning, run the task-requirement-engineer agent: understand → validate → feasibility → size → plan needs-answer, plan-review, or doing
needs-answer the human answers on the board, then writes answered: yes re-run this phase with the answers
plan-review the human writes approved: yes on the board doing
doing and after the task-workflow skill takes over unchanged —

Both human gates behave identically: the loop skips the task entirely until the confirming line is present. Neither ever times out into a guess.

PLAN_GATE decides whether plan-review happens at all:

PLAN_GATE plans you approve
always every plan
medium (default) MEDIUM and LARGE. SMALL is auto-approved.
small SMALL only (unusual; for trusting the RE on big work but not small)
never none — fully hands off

A skipped gate is not a skipped plan. plan.md is still written, still linked from the board, still the implementer's contract, and still reviewed as part of the PR.

A gate the knob skips is recorded as an approval, not as an absence of one:

      plan: workflow/TASK-31-reply-digest/plan.md  (SMALL · 1 slice)
      approved: yes (auto — PLAN_GATE=medium, size SMALL)

The board still reads status: plan-review, so the next run picks it up through the ordinary claim rule and nothing special-cases the bypass. You can also see, six months later, that no human ever read that plan.

The questions protocol

RE questions go on the board, short, with a stated default. The human answers in place. Long-form context stays in workflow/<folder>/questions.md.

- [ ] TASK-31 Reply digest for operator — status: needs-answer
      Operator gets a daily email digest of new replies grouped by niche.
      workflow: TASK-31-reply-digest
      Q1: One email with niche sections, or one email per niche? (default: one email)
      A1:
      Q2: Include replies from suppressed businesses? (default: no)
      A2:
      answered:
  • Max five questions. One line each. Every question names its default, so the human can accept the lot without typing an answer.
  • needs-answer is human-gated. A driver skips the task until the human writes answered: yes — same shape as approved: yes, same meaning: your turn is over. A loop ticking every few minutes must never answer its own questions.
  • On resume: A<n>: <anything> wins and goes into plan.md verbatim. An A<n> the human left empty means use the stated default — record it in assumptions.md and continue. Answer the two that matter, ignore the rest.

Handing the plan back to the board

When RE finishes, the dispatcher writes to the board (RE never touches tasks.md):

  1. Add the tag from plan.md's ## Agent section — the line stops being untagged, and every downstream rule works as it always has.
  2. Add the pointer: workflow: TASK-31-reply-digest.
  3. Add the plan line: plan: workflow/TASK-31-reply-digest/plan.md (MEDIUM · 2 slices · 1 migration).
  4. Set the status per PLAN_GATE.
- [ ] TASK-31 [backend] Reply digest for operator — status: plan-review
      Operator gets a daily email digest of new replies grouped by niche.
      workflow: TASK-31-reply-digest
      plan: workflow/TASK-31-reply-digest/plan.md  (MEDIUM · 2 slices · 1 migration)
      approved:

When the work is bigger than one sitting

RE writes proposed child lines to workflow/<folder>/tasks.md. The dispatcher appends them to the board with real ids and depends: lines, then closes the parent:

- [x] TASK-31 Reply digest for operator — status: done
      decomposed into TASK-32, TASK-33
- [ ] TASK-32 [backend] Build digest query + BuildDigestService with tests — status: todo
      plan: workflow/TASK-31-reply-digest/plan.md
- [ ] TASK-33 [backend] Schedule SendDigestJob daily 07:00 with tests — status: todo
      depends: TASK-32
      plan: workflow/TASK-31-reply-digest/plan.md

Children are tagged, so they skip intake entirely and share the parent's plan.

The export rule

workflow/ is gitignored. Anything that must outlive the task has to leave it before the task can reach done:

must land in what
the PR body the original requirement text, the plan, and the assumptions
tasks/architecture.md architecture-delta.md applied — LARGE only, inside the worktree
tasks/system_model.md system-model-delta.md applied — LARGE only, inside the worktree
the project's manual test cases updated whenever operator-visible behavior changed
docs/conventions.md or CLAUDE.md only if the task established a new project rule

A task must not be set to done while a durable artifact exists only inside workflow/. That single rule is what makes gitignoring the folder safe.

Absolute paths in worktrees

workflow/ is gitignored, so it does not exist inside a task worktree. An implementation agent running in ../<repo>-worktrees/TASK-31 that reads workflow/TASK-31-reply-digest/plan.md finds nothing — and will happily invent a plan instead of failing.

Every prompt handed to an implementation or review agent passes the plan by absolute path:

/Users/you/code/my-store/workflow/TASK-31-reply-digest/plan.md

Working directory is the worktree. Plan reads come from the main checkout.

Cleanup

A task reaching done keeps its workflow folder until the PR is merged (the review rounds need it). After that it is garbage. Deleting workflow/TASK-31-reply-digest/ costs nothing — the PR holds everything durable.