This skill should be used when the user asks about "how sprints work", "sprint phases", "iteration workflow", "convergent development", "sprint lifecycle", "when to use sprints", or wants to understand the sprint execution model and its convergent diffusion approach.
This skill inherits all available tools. When active, it can use any tool Claude has access to.
Sprint implements a convergent development model where autonomous agents iteratively refine implementations until specifications are satisfied. This skill covers the execution lifecycle, phase transitions, and iteration patterns.
Traditional AI-assisted development suffers from context bloat - each iteration adds more information, compounding errors and noise. Sprint reverses this pattern:
The metaphor is diffusion models in reverse - instead of adding noise to generate, remove noise to refine.
A sprint executes through 6 distinct phases:
Parse the sprint directory and prepare context:
.claude/sprint/[N]/)specs.md for user requirementsstatus.md if resumingThe project-architect agent analyzes requirements:
project-map.md for architecture contextproject-goals.md for business objectivesapi-contract.md, backend-specs.md, etc.)Spawn implementation agents in parallel:
python-dev for Python/FastAPI backendnextjs-dev for Next.js frontendcicd-agent for CI/CD pipelinesallpurpose-agent for any other technologyExecute testing agents:
qa-test-agent runs first (API and unit tests)ui-test-agent runs after (browser-based E2E tests)Architect reviews all reports:
status.md with current stateSprint completion:
status.md summaryWhen running /sprint on an existing sprint:
If status.md shows COMPLETE:
If status.md shows IN PROGRESS:
This ensures the user always knows where they are and what options they have.
The sprint cycles between phases 1-4 until complete:
┌─────────────────────────────────────────┐
│ │
│ ┌──────────┐ ┌──────────────────┐ │
│ │ Phase 1 │───▶│ Phase 2 │ │
│ │ Planning │ │ Implementation │ │
│ └──────────┘ └────────┬─────────┘ │
│ ▲ │ │
│ │ ▼ │
│ ┌────┴─────┐ ┌──────────────────┐ │
│ │ Phase 4 │◀───│ Phase 3 │ │
│ │ Review │ │ Testing │ │
│ └──────────┘ └──────────────────┘ │
│ │
└─────────────────────────────────────────┘
│
▼ (after max 5 iterations or success)
┌──────────┐
│ Phase 5 │
│ Finalize │
└──────────┘
Maximum 5 iterations: The system pauses after 5 cycles to prevent infinite loops. User intervention may be needed for complex blockers.
Sprints are ideal for:
Sprints are overkill for:
| File | Purpose |
|---|---|
specs.md | Requirements, scope, testing config |
project-goals.md | Business vision and objectives |
| File | Purpose |
|---|---|
status.md | Current sprint state (architect maintains) |
project-map.md | Technical architecture (architect maintains) |
api-contract.md | Shared interface between agents |
*-specs.md | Agent-specific implementation guidance |
| File | Purpose |
|---|---|
*-report-[N].md | Agent reports per iteration |
After each iteration, the architect:
This prevents context bloat and focuses agents on actual gaps.
status.md is rewritten each iteration, not appended:
Agents return machine-parseable reports:
There are two ways to do manual testing:
Set UI Testing Mode: manual in your specs.md:
## Testing
- UI Testing: required
- UI Testing Mode: manual
When the architect requests UI testing:
Use /sprint:test for quick testing outside of sprints:
.claude/sprint/[N]/manual-test-report.mdReports feed into sprints: When you run /sprint, the architect reads your manual test report and prioritizes fixing the issues you discovered.
Use manual testing for:
For more details, see the full command and agent documentation in the plugin's commands/ and agents/ directories.