| name | workflow-bootstrap |
| description | Create a new project from a prompt — drive a greenfield idea through the full 9-phase lifecycle (BA, design, UI, plan, code, review, test, fixbug, deploy), step-gated and persona-aware. Use when: starting a greenfield project, an empty repo and the user asked to build something, scaffolding a new app/service from a prompt, or workflow-intake routed case 1. |
Greenfield request: $ARGUMENTS
One-Liner
Setup runs once and per-US execution repeats: model project setup as F00 child-tasks (Stage A,
once) and every User Story as a workflow-feature run (Stage B, repeated), so any interrupt is
resumable from the active feature — never from chat.
Turn a prompt into a working, verified project. Reached from workflow-intake case 1, or directly
when the repo is empty and the user asked to build something.
Stage A (ONCE): F00-T1 BA → F00-T2 design → F00-T3 UI → F00-T4 plan+skeleton → close F00
Stage B (LOOP): for each backlog US → workflow-feature (code→review→test→verify→ship)
Pre-flight — seed the setup as a parent + child-tasks
Gate: persona/autonomy/collab are set, a session is active, and F00 + F00-T1…T4 exist
(parent still not-started).
Stage A — project setup (F00 child-tasks, runs ONCE)
Run the children one at a time (WIP=1), each as a full mini-loop —
harness plan F00-Tn → start → produce artifact → review by a design-derived role
(scripts/role-resolver.sh --phase design → an Architect; check-code-review for code artifacts) →
harness verify F00-Tn —
with a step-gate (../../resources/step-gate.md) between them and the
task-state file updated per child. Persona (../../resources/persona-mode.md)
sets depth/language.
| Child-task | Skill | Output artifact |
|---|
| F00-T1 ① Business analysis | plan-ba-analysis → plan-us-backlog | docs/requirements.md + seeded US backlog (harness add F01…) |
| F00-T2 ② System design | design-architecture (+ design-stack-research), design-database, design-api, design-detailed | docs/design/* incl. stack-versions.md |
| F00-T3 ③ UI design | design-ui, design-ux-flow | docs/design/ui/* |
| F00-T4 ④ Plan tasks | plan-tasks, plan-skeleton | child-tasks for the US + frozen test-strategy + green skeleton under source_dir/ (named after the system, not the repo root) |
Splitting Stage A into child-tasks gives design and UI their own ## Review evidence and finer
resume granularity — see plan-tasks (it decomposes Stage A the same way it
decomposes a US).
Close Stage A: once F00-T1…T4 are all passing (nothing in_progress), ./harness start F00
→ write docs/design-docs/F00/evidence.md (summary + links to the child evidence) → ./harness verify F00
→ passing. Only then start Stage B (this keeps WIP=1: the parent is started last, after its children).
Gate: F00 is passing (requirements + design + UI + seeded US backlog + green skeleton all
exist and verified), so Stage B has a working frame to build on.
Stage B — per-US execution (repeated, via workflow-feature)
For each backlog US, run phases ⑤–⑨ through workflow-feature:
harness plan <id> → start → ⑤ code (dev-*) → ⑥ review (check-review-loop:
independent role-based review — scripts/role-resolver.sh --phase review + harness review open/record provenance — + fix loop, check-refactor) →
⑦ test (check-test-strategy → test-*) → ⑧ fix (check-qa) → harness verify <id> →
⑨ ship (ship-commit-msg/ship-pr-create; deploy = always-stop). In solo you may auto-chain USs
(Non-Technical default); in team each assignee holds one US — see
workflow-team.
Gate: every targeted backlog US is passing (review + required tests in its evidence), or the
remaining ones are explicitly deferred with the user.
Resumability (why the two stages matter)
After any interrupt, ./harness resume reports the active feature + its task-state + next step:
- mid-Stage-A → the active F00-Tn child is
in_progress, its task-state says what is next (the
parent F00 stays not-started until its children pass).
- mid-Stage-B → the current US is
in_progress with its own task-state.
No bootstrap progress lives only in chat — it is always recoverable from features.json + .harness/tasks/.
Persona behavior (per phase)
- Non-Technical — Stage A asks requirement-level questions in plain language and picks sensible
technical defaults (logged); auto-advance; review/test still run fully but report in plain terms.
- Developer — Stage A is exhaustive (acceptance, schema, contracts, NFRs, test strategy); gated
after every phase; trade-offs surfaced for approval.
Cross-cutting
- Step-gate between every phase; always-stop before deploy and any irreversible action
(../../resources/autonomy-mode.md).
- Caveman (on by default) — apply opt-caveman to every inter-agent
prompt and subagent result across Stage A and Stage B, and write long artifacts to files
(../../resources/token-budget.md).
- Verify before done — no US is
passing without review + required tests recorded in evidence.
- Task-state — after
./harness start <id> expand .harness/tasks/<id>.md to the full template;
tick [ ]→[/]→[x] and update Next step at each phase boundary
(../../resources/task-state.md).