Goal and Epic Craft
Three workflows over one rubric, references/STANDARD.md:
craft (rough intent in, bounded goal out), grade (score an existing
goal), split (bounded epic into vertical slices). Read the standard
first; it defines every term below. All three take and produce plain
markdown. Filing into a specific tracker is an optional integration —
references/work-integration.md.
Craft
- Take the user's words as given. Keep the original phrasing; it becomes
the heart of
## Description. No metadata questions yet.
- Recon: facts are your job. Skim whatever the intent names — files, repos,
trackers — and check for neighboring or overlapping work; use
agents/recon.md for anything slow. The user answers
decisions; the environment answers facts.
- Interview in frontier rounds. Each round, ask every question from the
standard's ready test whose prerequisites are settled: numbered, with
your recommended answer, at most 5 per round; a typical session finishes
in 2. The user may leave any decision to the executor — record it in
## Notes; that counts as answered. Stop when the ready test drains.
- Draft the full document to the standard; title short, meaning
front-loaded. Show it, take one confirm round, and hand it over. Run
bin/goal-epic-craft DRAFT.md and repair what it reports before showing.
Grade
- Load the document (pasted, file, or tracker item). Missing input: ask —
never guess which item is meant.
- Run
bin/goal-epic-craft FILE for the mechanical evidence, then judge
the rest in the standard's order: four properties, ready test, failure
modes, fewer-sharper caps, epic check. agents/grader.md
runs this as an independent pass.
- Report: verdict first (ready to hand off or blocked — N gaps),
then at most 5 gaps ranked by how badly each would misdirect an executor
— failure-mode name, location, and the question whose answer cures it.
Praise nothing; change nothing yourself.
Split
- Gate: run the grade workflow silently on the epic. Blocked input stops
here — report the gaps and offer the craft workflow. Splitting an
unbounded epic multiplies its gaps into every child.
- Propose 3–7 children per the standard's epic extras: vertical, the first
the walking skeleton; title plus one goal sentence ending in a pointer to
the parent's purpose; 1–3 acceptance criteria each with a verification
method; an ordering between siblings only when one truly cannot start
first — a dependency chain is a horizontal split wearing a costume.
- One confirm round. The deliverable is one block per child under the
parent's title.
Whatever the workflow: the human promotes, schedules, or delegates the
result; never do that on their behalf.
1---2name: goal-epic-craft3description: Turn rough intent into a bounded goal or epic through a short interview, grade an existing goal against a written standard, or split a bounded epic into vertical slices. Use when a brief is too vague to execute, when asked to shape, bound, or scope a goal, epic, or project brief, or before delegating work to an autonomous agent.4---56# Goal and Epic Craft78Three workflows over one rubric, [references/STANDARD.md](references/STANDARD.md):9**craft** (rough intent in, bounded goal out), **grade** (score an existing10goal), **split** (bounded epic into vertical slices). Read the standard11first; it defines every term below. All three take and produce plain12markdown. Filing into a specific tracker is an optional integration —13[references/work-integration.md](references/work-integration.md).1415## Craft16171. Take the user's words as given. Keep the original phrasing; it becomes18 the heart of `## Description`. No metadata questions yet.192. Recon: facts are your job. Skim whatever the intent names — files, repos,20 trackers — and check for neighboring or overlapping work; use21 [agents/recon.md](agents/recon.md) for anything slow. The user answers22 decisions; the environment answers facts.233. Interview in frontier rounds. Each round, ask every question from the24 standard's ready test whose prerequisites are settled: numbered, with25 your recommended answer, at most 5 per round; a typical session finishes26 in 2. The user may leave any decision to the executor — record it in27 `## Notes`; that counts as answered. Stop when the ready test drains.284. Draft the full document to the standard; title short, meaning29 front-loaded. Show it, take one confirm round, and hand it over. Run30 `bin/goal-epic-craft DRAFT.md` and repair what it reports before showing.3132## Grade33341. Load the document (pasted, file, or tracker item). Missing input: ask —35 never guess which item is meant.362. Run `bin/goal-epic-craft FILE` for the mechanical evidence, then judge37 the rest in the standard's order: four properties, ready test, failure38 modes, fewer-sharper caps, epic check. [agents/grader.md](agents/grader.md)39 runs this as an independent pass.403. Report: verdict first (**ready to hand off** or **blocked — N gaps**),41 then at most 5 gaps ranked by how badly each would misdirect an executor42 — failure-mode name, location, and the question whose answer cures it.43 Praise nothing; change nothing yourself.4445## Split46471. Gate: run the grade workflow silently on the epic. Blocked input stops48 here — report the gaps and offer the craft workflow. Splitting an49 unbounded epic multiplies its gaps into every child.502. Propose 3–7 children per the standard's epic extras: vertical, the first51 the walking skeleton; title plus one goal sentence ending in a pointer to52 the parent's purpose; 1–3 acceptance criteria each with a verification53 method; an ordering between siblings only when one truly cannot start54 first — a dependency chain is a horizontal split wearing a costume.553. One confirm round. The deliverable is one block per child under the56 parent's title.5758Whatever the workflow: the human promotes, schedules, or delegates the59result; never do that on their behalf.
batteryshark