skills/agent-goal-skill/examples/multi-day-sprint.md

Example — Multi-day sprint

User input

/agent-goal-skill drain the auth-rewrite backlog. 11 PRDs. push to dev branch. no main.

Shaped output

```
/goal

# Goal
Drive all 11 PRDs in the auth-rewrite backlog from todo → done, with atomic commits per PRD on a feature branch, no pushes to main.

# Context
- Backlog source of truth: <TODO: kanban path or ticket-tracker query>
- 11 PRDs identified — each has acceptance criteria in its body
- Working branch: feat/auth-rewrite (create off main if not present)
- Sprint state tracking: {repo}/.sprint-state.json (per-PRD progress)
- Each PRD must pass the project's pipeline gates before commit
- Token budget across all 11: ~7M total (≈650K per PRD on average)

# Constraints
- One PRD per commit. Never merge two PRDs in one commit.
- NEVER push to main. NEVER force-push. Only push to feat/auth-rewrite.
- Skip PRDs tagged blocked, needs-design, manual-review — surface those in the final report
- Pipeline gates mandatory per PRD: lint, typecheck, tests, gap analysis. Never SKIP_PIPELINE_CHECK.
- If a PRD is missing acceptance criteria, mark it blocked and move on — do NOT invent criteria
- Final handoff: write {repo}/.claude/handoff-{YYYY-MM-DD}.md with per-PRD status + commit SHAs

# Done when
1. `kanban list --status "todo OR doing"` returns 0 unblocked actionable PRDs
2. Every done PRD has a commit SHA in its body
3. `git log --oneline feat/auth-rewrite ^main` shows ≤ 11 atomic commits (one per shipped PRD)
4. Final handoff written to {repo}/handoff-YYYY-MM-DD.md with per-PRD status table

# On block
FAIL FAST. If a PRD lacks acceptance criteria, mark blocked with reason, move to next. FAIL FAST only on infrastructure breaks (pipeline gate scripts missing, test runner broken, etc). Append all open questions to a pending-questions file.

Token budget: 7M
```

**Sandbox:** `workspace-write`
**Reasoning effort:** `xhigh`
**Caveats:** Backlog source TODO is required — confirm path before dispatching, otherwise the agent has no work queue to drain.

What the shaper did

• Sized token budget to sprint scope (7M across 11 PRDs)
• Set reasoning to xhigh because of multi-hour autonomy
• Anti-criteria explicit: NEVER push to main, NEVER force-push, NEVER SKIP_PIPELINE_CHECK
• Per-PRD atomicity rule (one commit per PRD)
• Defined what to do with broken PRDs (mark blocked, move on) without breaking the loop
• Required structured handoff at the end so the next session can pick up cleanly