with one click
implement
// Use when executing a feature plan with wave-based parallel agents. Triggers on "implement this plan", "start building", or when a plan exists in docs/plans/ and user wants to begin coding.
// Use when executing a feature plan with wave-based parallel agents. Triggers on "implement this plan", "start building", or when a plan exists in docs/plans/ and user wants to begin coding.
[HINT] Download the complete skill directory including SKILL.md and all related files
| name | implement |
| description | Use when executing a feature plan with wave-based parallel agents. Triggers on "implement this plan", "start building", or when a plan exists in docs/plans/ and user wants to begin coding. |
Execute a wave-structured plan by dispatching parallel agents. Each agent implements its assigned files independently using the plan's specification. Orchestrator stays lean—just coordinates and collects results.
The orchestrator reads ONLY the plan document (once, at start) and build error output (after failed builds). Never read source files (.cpp, .h, .fs, .glsl) or anything in src/ or shaders/. The plan has the complete spec — agents read codebase files with fresh context.
1. Load plan
2. Parse waves and tasks
3. For each wave:
a. Dispatch agents in parallel (one Task call per task)
b. Wait for all to complete
c. Build to verify wave
4. Code review
5. Final verification
6. Commit
Goal: Parse the plan document (this is the ONLY file you read before dispatching agents)
Actions:
Goal: Isolate work
Actions:
git checkout -b <kebab-case-from-filename>git checkout <branch-name>Goal: Run tasks with maximum parallelism
For each wave in order:
The plan contains the complete specification — file names, field names, line numbers, exact changes. Agents read their own target files with fresh context. The orchestrator pastes the plan's task section into the prompt verbatim and dispatches. Do NOT pre-read files to "enrich" agent prompts.
If a task's Do section references an external document for algorithm content (e.g., "from the research doc," "as specified in docs/research/"): STOP. Tell user: "Task [name] references [doc] for algorithm details — the plan should inline these. Run /feature-plan to fix." Do not proceed. Do not summarize the referenced document.
For each task in the wave, build a prompt containing:
# Task: [Task Name]
## Your Assignment
Implement the following files: [file list]
## Specification
[PASTE FULL SPEC SECTION FROM PLAN]
## Task Instructions
[PASTE BUILD SECTION FROM THIS TASK]
## Verification
[PASTE VERIFY SECTION FROM THIS TASK]
## Rules
- Read the files you need to modify BEFORE making changes
- Implement ONLY the files listed above
- Follow the specification exactly
- Use the codebase conventions from CLAUDE.md
- If touching UI code, invoke the `/ui-guide` skill
- Do NOT modify files outside your assignment
- Do NOT run cmake or any build commands (orchestrator handles builds)
- Do NOT run the application
- Report completion with list of files modified
CRITICAL: Spawn all tasks in a wave with a SINGLE message containing multiple Task calls:
Task(prompt="[task 1 prompt]", subagent_type="general-purpose", description="Implement [file1]")
Task(prompt="[task 2 prompt]", subagent_type="general-purpose", description="Implement [file2]")
Task(prompt="[task 3 prompt]", subagent_type="general-purpose", description="Implement [file3]")
All agents run in parallel. Tool blocks until all complete.
Do NOT:
After all agents return:
After each wave:
cmake.exe --build build
Goal: Catch correctness and consistency issues before the user runs the feature
Actions:
git diff main...HEAD --stat and capture the outputcode-reviewer agent covering all focus areasgit diff main...HEAD --stat output, focus on both correctness and consistency (code should read like one author, not N agents). Include in prompt: the stat output and instruction "Use Read to inspect the changed files listed above. Do NOT run git or bash commands."Goal: Confirm feature works
Actions:
Goal: Record the work
Actions:
git add .)Each task OWNS its files exclusively. No two tasks in the same wave modify the same file.
If a plan has file overlap within a wave: the plan is wrong. Stop and report to user.
Each agent receives:
Agents do NOT:
Agents may discover work not in the plan. Apply these rules:
Auto-fix (no permission needed):
Ask user:
Track all deviations for the commit message.
Agent reports error:
Build fails after wave:
Plan structure:
Execution:
Wave 1: Dispatch 1 agent → config header created → build succeeds
Wave 2: Dispatch 7 agents simultaneously:
- Agent A: effect_config.h modifications
- Agent B: shader creation
- Agent C: post_effect.h/cpp
- Agent D: shader_setup.h/cpp
- Agent E: imgui_effects*.cpp
- Agent F: preset.cpp
- Agent G: param_registry.cpp
All 7 complete → build → final verification → commit
Total: 2 build cycles, not 8.
| Thought | Reality |
|---|---|
| "Let me read a source file first" | The plan has the spec. Agents read codebase files. Dispatch. |
| "I'll just do these tasks myself" | Dispatch agents. Fresh context = better quality. |
| "I'll run them one at a time to be safe" | Parallel is safe if no file overlap. Check the plan. |
| "The build failed, I'll fix it later" | Fix now. Don't proceed with broken build. |
| "I'll skip the small tasks" | All tasks matter. Dispatch them all. |
| "This agent's work looks wrong" | Check against spec. If spec is right, agent is probably right. |
| "The task says 'from research doc' — I'll include the algorithm" | STOP. The plan is incomplete. Tell the user to fix it. |