| name | implementation |
| description | Execute a technical plan task by task, following the phased structure.
v1.0: Hooks handle verification automatically - no manual reminders needed.
Use when a technical plan exists and it's time to write code.
|
Implementation Skill
Goal
Execute a technical plan systematically, following established phases, completing tasks in order.
Activation Triggers
- User says "implement", "build", "execute the plan", or "start coding"
- A technical plan exists and it is time to write code
- User runs
/omg:implement
Hook Integration
Hooks handle verification automatically:
AfterTool hook runs typecheck/lint after code changes
phase-gate hook detects phase completion
ralph-retry hook handles persistence mode
You no longer need to manually remember verification steps.
Process
1. Load Plan & Context
find . -name "tasks_*.md" -o -name "plan.md" 2>/dev/null | head -5
cat GEMINI.md 2>/dev/null | head -50
cat conductor/tech-stack.md 2>/dev/null | head -30
2. Identify Current State
Parse the plan to find:
- Which tasks are already complete
[x]
- Which task is next
[ ]
- Which phase we're in
3. Execute Sequentially
Work through tasks in order within each phase.
Execution Rules
Phase Order is Mandatory
- Complete Phase 1 (Data Layer) before Phase 2
- Complete Phase 2 (Backend) before Phase 3
- Complete Phase 3 (Frontend) before Phase 4
- Never skip ahead
Task Order Within Phases
- Execute tasks in numerical order (1.0, 1.1, 1.2...)
- Complete all sub-tasks before marking parent complete
- Don't start new parent task until previous is done
Automatic Verification (via Hooks)
After you modify code files, the AfterTool hook automatically:
- Runs typecheck (if TypeScript)
- Runs lint (if configured)
- Injects any errors into context
If errors appear, fix them before proceeding to the next task.
Phase Gates (via Hooks)
When you complete a phase, the phase-gate hook:
ask_user available (v0.30.0+): Instructs you to call ask_user with a yes/no verification question. Use this to confirm with the developer before proceeding to the next phase.
- Advisory mode (default): Shows a message suggesting verification
- Strict mode: Requires explicit user confirmation
The hook detects phase completion from your response and handles the gate automatically.
Verification Checkpoints
When completing a phase, use ask_user to confirm with the developer before proceeding:
ask_user({ question: "Phase 'Data Layer' is complete. Have you verified all tasks?", question_type: "yes_no" })
If ask_user is not available (headless mode, older CLI), the phase-gate hook falls back to prompt-based verification.
Implementation Standards
Before Writing Code
cat src/[similar-file].ts | head -100
grep -r "export function" src/utils/ | head -20
While Writing Code
DO:
- Follow existing code style
- Use established patterns from codebase
- Include appropriate error handling
- Add meaningful variable/function names
DON'T:
- Introduce new patterns without reason
- Skip error handling
- Leave TODO comments without tracking
After Each Task
- The hook verifies automatically
- If errors, fix them (errors appear in context)
- Update task status to
[x] in plan
- Move to next task
Output Formats
After Completing Each Task
## ✅ Completed: [Task Number] [Task Name]
### Changes Made
- `path/to/file.ts`: [What was done]
- `path/to/new-file.ts`: Created - [purpose]
### Next Task
[Task number and name]
At Phase Boundaries
The phase-gate hook will prompt for verification. Respond with:
## 🔍 Phase [N] Complete: [Phase Name]
### Summary
[What was accomplished in this phase]
### Files Changed
[List of files]
### Ready to proceed to Phase [N+1]?
After Completing All Phases
## 🎉 Implementation Complete: [Feature Name]
### Summary
[Overall summary of what was built]
### All Phases Completed
- [x] Phase 1: Data Layer & Types
- [x] Phase 2: Backend API
- [x] Phase 3: Frontend Implementation
- [x] Phase 4: Review & Finalize
### Next Steps
- Code review (optional)
- Documentation updates (if needed)
Error Handling
If a Task Fails
The AfterTool hook injects errors into context. When you see them:
- Read the error messages
- Identify the file and line
- Fix the specific issue
- Your next edit will be verified again
If Blocked by Missing Information
- Document what's needed
- Check if it's in the spec/PRD
- If not, ask the user
- Don't proceed with assumptions on critical decisions
Persistence Mode
If the user activates Ralph mode (includes "ralph" or "persistent" in prompt):
- The
ralph-retry hook activates
- On failure, it suggests alternative approaches
- Keeps retrying up to configured max
Plan Mode Compatibility
When plan mode is active, use it to review and decompose complex tasks before
implementing. Exit plan mode before making code changes.
Guidelines
- One task at a time: Focus, complete, move on
- Trust the hooks: Verification happens automatically
- Match the codebase: New code should look like it belongs
- Follow the plan: Don't improvise unless blocked
- Update status: Keep plan.md reflecting reality