SDD Tasks
The Tasks phase of SDD. Turn plan.md into a list of tasks you could hand to an
implementer as they are: each says what to build, where, and which acceptance criteria it must
satisfy. You plan and describe; you do not implement and you do not choose agents or models —
/sdd:implement hands each task to the project's implementation flow.
Steps
1. Load context
- the feature folder
specs/<current>/: from$ARGUMENTS, else the spec whose**Branch:**equalsgit branch --show-current(none → list the specs and ask); read itsplan.mdandspec.md(the AC ids) specs/capabilities.yml:projectsandactions.commit(step 5). Only the oldspecs/capabilities.mdexists → STOP: "Run/sdd:doctor initto migrate it."tasks.md, when it already exists (a re-run after the plan or the spec changed)
2. Decompose
Hierarchy: Epic → Task. The Epic is the feature, exactly one per /sdd:spec run (work that
feels like several Epics should have been separate specs — say so instead of inventing Epics).
A task is a ticket: one project, one coherent slice of behaviour, typically a few hours to a day — a whole screen with its state and data loading, a set of endpoints with their logic and storage, a migration, a CLI command. Not a layer, not a micro-step: no task for "the form", for "the tests" or for "the service" of one unit. Tests belong to the task that delivers the behaviour.
- Every spec AC is covered by at least one task; every task serves at least one AC.
- A contract shared between projects (schema, generated client, shared types, migration) is its own task in the project that owns it; consumers depend on it. Cross-project verification (e2e) is a task that depends on what it exercises.
depends_ononly where a task needs another's result or touches the same files. Never because of list order.- Do not copy the plan or the spec. Refer to AC ids and plan sections;
/sdd:implementinlines them into the brief when it hands the task over.
3. Epic id
The Epic gets an id that never changes: <LETTERS>-<4 hex>, for example ULOG-7f3a. Tasks keep
their T<n> headings; the full id of a task is <EPIC>-T<n> (ULOG-7f3a-T3). It is what
/sdd:implement puts in commit trailers and task branch names.
tasks.mdhas a**Epic:**line → keep it.- Otherwise build it: 2–5 capital letters from the spec slug (first letter of each word; one word →
its first 4 letters),
-, and 4 hex characters fromprintf '%s %s' "<spec branch>" "$(date +%s)" | git hash-object --stdin | cut -c1-4. Anotherspecs/*/tasks.mdalready has that id → hash again.
4. Write specs/<current>/tasks.md
# Tasks: <feature_title>
**Spec:** `specs/<slug>/spec.md` **Plan:** `specs/<slug>/plan.md`
**Epic:** <EPIC>
## Epic: <feature_title>
### T1: <title>
project: <name from `projects`; `root` in a single-app repo>
status: draft
depends_on: []
acceptance: AC1, AC3
plan: <plan sections this task implements>
scope:
- <concrete action>
- <concrete action>
files: <expected paths or globs, relative to the spec repo>
notes: <optional — a constraint or constitution rule that applies here>
scope is the list of actions, specific enough to start from (behaviours, states, edge cases from
the AC). acceptance is measurable via the spec's AC ids. Write depends_on: [] when the task
waits for nothing; omit notes when nothing applies.
Re-run (a tasks.md exists): update it, do not start over. Keep the **Epic:** id (an old
file without one gets one) and each existing task's id, status, branch: and pr: lines; a task that is in-progress, review or done keeps its scope unless the change
touches it — then set it back to draft and say so. New tasks take the next unused id; a task no
longer needed is removed only while draft, otherwise report it and leave it.
Add one line at the end: Order: T1 → T2 ∥ T3 → T4 (what can run once its dependencies are done).
5. Check, advance, commit, report
Fail and fix on: an AC with no task, a task with no AC, a cycle, a dependency on a missing task, two
independent tasks with overlapping files.
Set plan.md **Status:** → tasks-ready with Edit (unless already later).
Commit the documentation. The spec, plan and tasks are the whole documentation of the feature, and nothing earlier commits them:
git status --short -- specs/<slug>/is empty → nothing to commit.- Otherwise stage
specs/<slug>/only (nevergit add -A, never other paths) and commit it with theactions.commit(skill:/agent:/command:); unset →git add -- specs/<slug>/andgit commit. Messagedocs(<slug>): add spec, plan and tasks(a re-run:docs(<slug>): update tasks), trailerEpic: <EPIC>. Check it landed withgit log -1 --format='%(trailers:key=Epic,valueonly)'; empty (a skill rewrote the message) →git commit --amend --trailer 'Epic: <EPIC>'. - Not on the feature branch (
**Branch:**of the spec) → do not commit; say so.
Report the task count, projects, the commit and the Order: line. Next: /sdd:implement T1 (one task) or /sdd:implement all; best after
/clear — it re-reads tasks.md from disk.
Do not write code.