Spec-driven dev with five agents, and you in the loop.
Five role-based agents — PM, designer, monetization strategist, engineer, reviewer — hand off through reviewable Markdown artifacts. By default, two checkpoints and a fully autonomous loop.
Works in Claude Code · OpenCode · Copilot CLI · Cursor · Windsurf · Codex · and any AGENTS.md tool.
→ francescoesposito.dev/speckroo for the full docs
Claude Code — native plugin:
/plugin marketplace add phoinixi/speckroo
/plugin install speckroo@speckroo
Everything else — one command in your project:
npx github:phoinixi/speckroo setup <tool>
<tool> is opencode, copilot, cursor, windsurf, or codex.
1. Install and scaffold:
/plugin marketplace add phoinixi/speckroo
/plugin install speckroo@speckroo
/speckroo:init-framework
init-framework walks you through filling .framework/constitution.md with your project's principles, and seeds queue.md + skill.md.
2. Add feature ideas to the queue (.framework/queue.md):
- [ ] user authentication
- [ ] dashboard charts
- [ ] CSV export3. Run the loop:
/speckroo:loop
The agent picks "user authentication", runs the PM and designer, then stops:
Approve shape (spec.md + design.md) and continue to /plan? (yes / revise / stop)
Say yes → the engineer plans and stops again:
Approve plan and continue to /build? (yes / revise / stop)
Say yes → the engineer builds all tasks → the code reviewer writes review.md → "user authentication" is checked off → the loop picks "dashboard charts".
That's it. Two decisions per feature.
If you want to target a single feature without using the queue:
/speckroo:shape user authentication
Same inline approval flow, just for that one feature.
Default (3 steps + auto review, 2 checkpoints):
/shape <idea> → spec.md + design.md ↓ "yes"
/plan → plan.md + tasks.md ↓ "yes"
/build → code (all tasks, one summary)
/review → review.md (adversarial check, informational)
Each checkpoint is an inline "yes". /review chains automatically — no prompt needed.
Loop mode (queue-driven, autonomous between gates):
# add ideas to .framework/queue.md, then:
/loop → run next feature in the queue, stop at approval gates
/loop check → inspect queue state without running anything
Strict mode (granular, every phase explicit):
/discover <idea> → spec.md ↓ /approve spec
/design → design.md ↓ /approve design
/monetize → monetization.md ↓ /approve monetize (optional)
/plan → plan.md + tasks ↓ /approve plan
/build next → code ↺ one task at a time
/status shows where every feature stands.
/approve is the explicit fallback for approvals — the normal path is an inline "yes".
One canonical source in core/ — five persona bodies, eleven command bodies, one workflow contract — projected into each tool's config layout by the CLI. The Claude Code plugin is byte-checked against core/ by CI.
core/personas/ ← 5 role prompts, written once
core/commands/ ← 11 command bodies, written once
core/workflow.md ← gate rules + loop + skill contract, written once
bin/speckroo.mjs ← projects core/ into any tool
speckroo/ ← the Claude Code plugin
docs/ ← GitHub Pages site
AGENTS.md ← universal fallback (no install needed)
This repo is built using speckroo. The full artifact trail — spec.md → design.md → plan.md → tasks.md — lives in .framework/speckroo-docs/.
MIT · NOTICE · CONTRIBUTING