Delegation Planner Skill
Purpose: Route every task to the correct agent layer before execution begins. Wrong routing wastes context budget and produces inferior output.
When to Use
Use this skill when:
- An incoming task needs to be assigned to an agent
- A complex task needs to be decomposed across multiple agents
- A senior agent is being asked to do sub-agent work (or vice versa)
- Before dispatching any specialist agent
Agent Layer Routing Rules
Layer 1: Senior PM Agent
Route here when:
- User speaks in outcome language ("finished", "ready", "ship it")
- A Definition of Done needs to be established
- Proof artifacts need to be defined
- Milestone planning is needed
- Stakeholder communication is required
Do NOT route here:
- Implementation tasks (code, tests, deployments)
- Technical debugging
- Research tasks
Layer 2: Senior Orchestrator Agent
Route here when:
- Task spans multiple specialist domains
- Phase dependencies need to be enforced
- Evidence needs to be collected from multiple agents
- A completion claim needs to be blocked or approved
- Context budget management across agents is needed
Do NOT route here:
- Single-domain implementation
- Research-only tasks
- Simple file edits
Layer 3: Senior Specialist Agents
Route to Senior Engineering Agent when:
- Backend API, FastAPI routes, SQLAlchemy models, LangGraph agents
- Database migrations, schema changes
- Authentication, JWT, middleware
- Performance optimisation, infrastructure
Route to Senior UI/UX Agent when:
- React components, Next.js pages, Tailwind CSS
- Design system compliance, animations
- Responsive layout, accessibility
- Visual quality review
Route to Senior QA / Production Agent when:
- Test coverage (vitest, pytest, Playwright)
- CI/CD pipeline setup or fixes
- Deployment verification
- Monitoring and alerting setup
Route to Senior Research Agent when:
- Technology comparison or evaluation
- External documentation gathering
- Competitor analysis
- Best practice investigation
Route to Senior LMS Content Agent when:
- Educational content creation
- Learning module structure
- Content pipeline setup
Route to Senior Growth / Marketing Agent when:
- SEO optimisation
- GEO (Generative Engine Optimisation)
- Analytics setup
- Conversion optimisation
Layer 4: Sub-Agents
Route here when:
- Single file edit or creation
- Targeted grep/glob search
- Proof artifact collection (curl, test run)
- Data transformation on a specific file
- Any task completable in < 5 minutes with a single tool
Do NOT route here:
- Multi-file architectural changes
- Tasks requiring decision-making
- Tasks requiring context from multiple parts of the codebase
Procedure
Step 1: Classify the task
Ask: What kind of work is this?
- Outcome language / stakeholder → Layer 1
- Multi-agent coordination → Layer 2
- Domain-specific implementation → Layer 3 (which specialist?)
- Isolated, bounded task → Layer 4
Step 2: Check for decomposition
If the task touches multiple domains (e.g., "add a feature with API + UI + tests"):
- Decompose into sub-tasks per domain
- Identify dependencies (API must exist before UI can call it)
- Plan execution order
Step 3: Generate delegation plan
Produce the delegation plan table.
Output Format
DELEGATION PLAN
═══════════════════════════════════════════════════
Task: [original task description]
Decomposed: [yes / no]
ROUTING
─────────────────
| # | Sub-task | Agent | Layer | Evidence Required | Gate |
|---|---------|-------|-------|------------------|------|
| 1 | [task] | [agent] | [1-4] | [what to return] | [gate condition] |
| 2 | [task] | [agent] | [1-4] | [what to return] | [gate condition — depends on #1] |
...
DEPENDENCY ORDER
─────────────────
[If tasks have dependencies: show order and gate conditions]
Step 1: Task #[N] must complete before Task #[N+1] begins
Gate: [exact condition that unlocks next step]
CONTEXT BUDGET
─────────────────
Orchestrator: [estimated tokens for coordination]
Each specialist: [estimated tokens per agent]
Total estimate: [rough total]
═══════════════════════════════════════════════════
Validation Gates
Before finalising delegation plan:
- [ ] No Layer 1 agent assigned implementation work
- [ ] No Layer 4 sub-agent assigned architectural decisions
- [ ] Dependencies are correctly ordered (no circular deps)
- [ ] Evidence required is specific (not "show your work")
- [ ] Gate conditions are binary (pass/fail, not subjective)
Failure Modes
| Failure | Recovery | | ------------------------------------------------ | ----------------------------------------------------------------- | | Task is too vague to route | Ask: "What does success look like?" — activate outcome-translator | | Task spans all 4 layers | Decompose into phases, route each phase separately | | Wrong agent assigned (discovered mid-task) | Re-route immediately, don't continue in wrong layer | | Sub-agent scope creeps to architectural decision | Pause sub-agent, escalate to orchestrator |
Eval Examples
Good Example
Task: "Add a protected dashboard page with user stats"
Delegation Plan: | # | Sub-task | Agent | Layer | Evidence Required | |---|---------|-------|-------|-----------------| | 1 | API endpoint: GET /api/user/stats | Senior Engineering | 3A | curl output + type-check | | 2 | Dashboard page component | Senior UI/UX | 3B | Screenshot + 0 TS errors | | 3 | Auth middleware check | Senior Engineering | 3A | curl 401 without JWT | | 4 | E2E test for dashboard | Senior QA | 3C | Playwright test output |
Dependency order: 1 → 3 → 2 → 4
Bad Example (rejected)
Task: "Add dashboard" Response: "Sure, I'll add the dashboard." — No routing, no decomposition, no evidence plan.