en un clic
en un clic
UI/UX design intelligence with searchable database
Generate comprehensive implementation plans through systematic discovery, synthesis, verification, and decomposition into beads. Use when asked to plan a feature, create a roadmap, design an implementation approach, or decompose work into trackable issues. Do NOT use for simple one-step tasks, quick fixes, or when the user just wants to execute an existing plan — use the work skill instead.
Deep investigation mode. Gather context, analyze, synthesize recommendations without making code changes.
Fetch up-to-date library documentation via Context7 MCP. Use when working with external libraries, APIs, or frameworks.
Start interview-driven planning with Prometheus. Asks clarifying questions before generating implementation plan.
Diagnose and fix Maestro installation issues
| name | work |
| description | Execute a plan or direct task with worker delegation and verification. |
.maestro/plans//work fix the login bug/work --resumeYou are the orchestrator.
| Argument | Meaning |
|---|---|
<plan-name> | Load a plan by name or title substring |
--resume | Skip completed tasks (- [x]) |
| (none) | Auto-load if one plan; prompt if multiple |
| (description) | Planless mode — see reference/planless-flow.md |
Check .maestro/handoff/*.json for "status": "designing". If found, warn the user and wait for confirmation.
Scan .maestro/plans/*.md for available plans.
If <plan-name> argument given:
add, fix, create, update, implement, refactor, remove, change, move, build). If so → reference/planless-flow.mdPlan not found. Available: {list}If no argument:
No plans found. Run /design first, or /work <description>.Validate the plan has:
## Objective section (required)- [ ] checkbox (required)## Verification section (required)## Scope section (warn if missing, proceed)Present summary: title, objective excerpt, task count, scope.
Ask: "Execute this plan?" — wait for explicit confirmation. If cancelled, stop.
Write handoff file:
// .maestro/handoff/{plan-slug}.json
{
"topic": "{slug}",
"status": "executing",
"started": "{ISO timestamp}",
"plan": ".maestro/plans/{slug}.md"
}
Worktree isolation (optional) — ask user, then follow reference/worktree-isolation.md if chosen.
Load priority context from .maestro/notepad.md under ## Working Memory. Items tagged P0-P2 become hard constraints appended to every task prompt.
Discover skills → see reference/skill-injection.md. Build a registry for task prompt injection.
Work through each - [ ] checkbox in the plan, in dependency order.
For each task:
Use this template for every worker assignment:
## TASK
[Task title from the checkbox]
## EXPECTED OUTCOME
- [ ] Files created/modified: [paths from plan]
- [ ] Tests pass: [test command]
- [ ] No new errors
## CONTEXT
[Background, constraints, related files from plan]
## SKILL GUIDANCE
[Only if matching skills found — see reference/skill-injection.md]
## MUST DO
- [Explicit requirements from plan]
## MUST NOT DO
- [Explicit exclusions]
## PRIORITY CONTEXT
- [P0-P2 items from notepad, if any]
## KNOWLEDGE CAPTURE
- If you discover a non-obvious constraint, append to .maestro/notepad.md under ## Working Memory.
- Emit <remember category="learning">description</remember> for non-trivial patterns.
Spawn a worker subagent with the task prompt. Use whatever delegation tool your runtime provides (Task, handoff, subagent, etc.) — the point is isolated execution so you can verify independently.
Worker role selection:
| Role | Use for |
|---|---|
kraken | TDD, new features, multi-file changes |
spark | Quick fixes, single-file changes, config |
build-fixer | Build errors, lint failures, type errors |
If no subagent tool is available, do the work yourself inline.
After the worker finishes → see reference/verification.md for the full protocol.
Summary:
If passes → commit and mark done (4d).
If fails → retry up to 3 times:
build-fixer for persistent build/lint errorsoracle for diagnostic help on attempt 3git add <file1> <file2>feat(<plan-scope>): <short task description>
Plan: <plan-name>
Task: <task subject>
- [x] Task title <!-- commit: abc1234 -->
Never end a session with 0 commits. Each verified task gets its own commit.
Move to the next - [ ] checkbox. Repeat 4a-4d until all tasks are done.
--resume mode: Skip all - [x] checkboxes.
After all tasks are done, pass these gates before wrapping up:
## Verification section.## Security section → see reference/security.mdcritic agent for final review. If verdict is REVISE, send issues to workers and wait for fixes.All gates must pass before proceeding.
Extract wisdom → see reference/wisdom.md
Write learnings to .maestro/wisdom/{slug}.md.
Archive plan:
mkdir -p .maestro/archive/
mv .maestro/plans/{slug}.md .maestro/archive/{slug}.md
Update handoff:
{ "status": "archived", "completed": "{ISO timestamp}" }
Worktree cleanup (if used) → see reference/worktree-isolation.md
Append summary to notepad (.maestro/notepad.md under ## Working Memory):
- [{date}] [work:{slug}] Completed: {N}/{total} tasks. Files: {count}. Security: {pass|skip}.
Tell the user:
.maestro/archive/{slug}.mdSuggest: Run /review to verify results against acceptance criteria.
| Don't | Do instead |
|---|---|
| Edit files yourself | Delegate to workers |
| Skip verification | Verify every task before committing |
| Use one-line task prompts | Use the full delegation template |
| Skip wisdom extraction | Always write .maestro/wisdom/ file |
| Forget to commit | Commit after each verified task |
| Start without handoff file | Always write handoff in Step 3 |
| Topic | File |
|---|---|
| Verification protocol | reference/verification.md |
| Security review | reference/security.md |
| Wisdom extraction | reference/wisdom.md |
| Worktree isolation | reference/worktree-isolation.md |
| Skill injection | reference/skill-injection.md |
| Planless flow | reference/planless-flow.md |