execute

name: execute description: Execute implementation plans task by task with verification. Use after /plan to implement the feature. Follows each task exactly, runs verifications, commits frequently. Triggers on "execute the plan", "implement the plan", "run the plan", or after /plan approval.

Execute: Plan Implementation

Execute implementation plans task by task. Follow each step exactly, run all verifications, commit after each task.

Announce at start: "I'm using the execute skill to implement this plan."

Prerequisites

Before executing:

1. An implementation plan exists (from /plan)
2. The plan has been approved
3. You're on the correct branch

If no plan exists, suggest running /plan first.

Workflow

┌──────────────┐    ┌──────────────┐    ┌──────────────┐    ┌──────────────┐
│ 1. Load Plan │───▶│ 2. Execute   │───▶│ 3. Verify    │───▶│ 4. Complete  │
│ & Review     │    │ Each Task    │    │ & Commit     │    │ & Review     │
└──────────────┘    └──────────────┘    └──────────────┘    └──────────────┘

Step 1: Load Plan & Review

1.1 Load the Plan

Read .claude/scratch/plans/[feature-name]-plan.md.

1.2 Critical Review

Before starting, check:

• Plan makes sense
• No obvious gaps
• Dependencies are clear
• On correct branch

If concerns exist, raise them before proceeding.

1.3 Create Progress Tracker

Track progress in .claude/scratch/tasks/todo.md:

## [Feature] Implementation

- [ ] Task 1: [Title]
- [ ] Task 2: [Title]
...

Carbon-Specific Skills to Load

When executing tasks, load the appropriate Carbon skill for guidance:

For migrations: Read .claude/rules/workflow-database-migration.md before writing SQL to ensure:

• id() function for primary keys
• companyId with composite primary key
• Standardized RLS policies (SELECT, INSERT, UPDATE, DELETE)
• Audit columns (createdBy, createdAt, updatedBy, updatedAt)

For forms: Invoke /forms to get ValidatedForm patterns, zod conventions, and action handlers.

Step 2: Execute Each Task

For each task:

2.1 Mark In Progress

Update .claude/scratch/tasks/todo.md:

- [x] Task 1: Create migration ✓
- [ ] Task 2: Add types (in progress)
- [ ] Task 3: Service layer

2.2 Follow Steps Exactly

The plan has bite-sized steps. Follow them exactly:

• Use the exact file paths
• Use the exact code
• Run the exact commands

2.3 Run Verifications

Every task has verification steps. Run them all:

• Tests must pass
• Commands must produce expected output
• TypeScript must compile

2.4 Commit

After task passes verification:

git add [specific files] && git commit -m "[type](module): [description]"

2.5 Mark Complete

Update tracker and move to next task.

Step 3: Handle Blockers

When to Stop

Stop immediately when:

• Test fails and you can't quickly fix it
• Plan has a gap (missing step or unclear instruction)
• Verification produces unexpected output
• You need to make a decision not covered by the plan

What to Do

1. Don't guess — Stop and ask
2. Document the blocker — What happened, what you tried
3. Propose options — If you have ideas, share them

After Blocker Resolution

Return to the task and continue.

Step 4: Complete & Review

4.1 Final Verification

After all tasks:

• All tests pass: pnpm test
• TypeScript compiles: pnpm typecheck
• Lint passes: pnpm lint

4.2 Feature Verification

Use /verify to test the feature in the browser:

• Happy path works
• Edge cases handled
• No regressions

4.3 Self-Review

Use /self-review to check your work before declaring done.

4.4 Report Completion

## Implementation Complete

**Branch:** feature/[name]
**Commits:** N commits
**Tests:** X new tests, all passing

**Summary:**
- [What was built]
- [Key decisions made during implementation]
- [Any deviations from plan and why]

**Ready for:** PR creation / further review

Subagent Strategy

For plans with independent tasks, consider dispatching subagents:

## Parallelizable Tasks

Tasks 3, 4, 5 are independent (all depend only on Task 2).
Dispatch 3 subagents in parallel:
- Subagent A: Task 3
- Subagent B: Task 4  
- Subagent C: Task 5

Use the Agent tool with clear context for each subagent.

Output

Next Steps

After execution:

• Create PR (if ready)
• Run /self-review
• Use /verify for browser testing