orchestrating-subagents
Design subagent orchestration plans for multi-step work. Use when a task needs decomposition across multiple specialized agents, model selection, tool assignment, skill selection, handoff contracts, s
SKILL.mdskills/orchestrating-subagents/SKILL.md
GitHubOrchestrating Subagents
Design the orchestration plan for subagent work.
Agent roles
Orchestrator
- Agent/model: Opus
- Role: planner, coordinator, reviewer, and synthesis owner
- Responsibility: decompose the task, assign subagents, define handoffs, detect gaps, and consolidate outputs
Default subagents
openai-codex/gpt-5.4- Use for coding tasks
- Use for technical research tasks
- Use for repo analysis, debugging plans, implementation planning, and validation design
gemini-3-pro-image-preview- Use for image creation tasks
- Use for visual concept generation, mockups, ad/image directions, and image-edit planning
Orchestration procedure
- Restate the target outcome in one sentence.
- Split the work into bounded subtasks with explicit deliverables.
- Mark each subtask as one of: coding, research, image creation, review, or coordination.
- Assign the best-fit agent/model.
- Assign only the tools and skills needed for that subtask.
- Define the artifact each subagent must return.
- Define dependency order and what can run in parallel.
- Define review gates for the orchestrator.
Tool assignment guidance
For coding tasks with openai-codex/gpt-5.4
Prefer tools such as:
read,write,edit,apply_patchexecandprocesswhen execution is allowed by the parent taskbrowserfor web UI inspection when neededweb_fetchonly for documentation or primary-source lookupdiffsfor real diffs
Prefer skills such as:
- the most task-specific repo/domain skill available
coding-agentwhen delegation to a coding harness is part of the planfrontend-designor other domain skills when relevant
For research tasks with openai-codex/gpt-5.4
Prefer tools such as:
web_fetchbrowserreadmemory_searchandmemory_getwhen prior decisions matter
Prefer skills such as:
- the most specific analysis or domain skill available
clawhubonly as a recommendation path when identifying missing skills
For image creation tasks with gemini-3-pro-image-preview
Prefer tools such as:
image_generateimagereadfor brand/context docs
Prefer skills such as:
ad-designerfrontend-design- other image/design skills matching the request
Missing capability rule
When a required tool or skill is missing:
- name the missing capability clearly
- check for a close existing substitute
- if no substitute fits, recommend fetching a skill from ClawHub
- if no suitable skill is likely available, recommend creating a new skill
- note the effect on sequencing and risk
Output format
Return these sections in this order:
- Outcome
- Orchestrator
- Subtasks
- Subagent assignments
- Tools per subagent
- Skills per subagent
- Handoffs and artifacts
- Parallelization
- Gaps
- Review gates
Keep assignments concrete. Avoid vague labels like "general agent" or "use standard tools".