From code-flow
Use when deciding how to structure work across agents — choosing between agent teams, subagents, or serial execution, and writing effective agent prompts.
How this skill is triggered — by the user, by Claude, or both
Slash command
/code-flow:agent-orchestrationThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Guide for deciding when and how to delegate work to other agents, and how to communicate with them effectively. The goal is maximum parallelism with minimum wasted tokens.
Guide for deciding when and how to delegate work to other agents, and how to communicate with them effectively. The goal is maximum parallelism with minimum wasted tokens.
Use when tasks need coordination — shared task lists, message passing, dynamic work assignment.
Team lead stays lightweight. The lead orchestrates — it creates tasks, assigns work, resolves blockers, and reviews. It should not be doing implementation work itself unless the team is idle.
Use when tasks are fully independent — no shared state, no coordination needed.
Use when work is tightly coupled or trivial.
The most important skill in orchestration is writing good prompts for the agents you spawn.
Agents will read files themselves. Don't paste file contents into prompts — point them to the files.
Do:
Modify
src/auth/login.tsto add rate limiting to thehandleLoginfunction. Follow the same pattern used insrc/auth/register.ts.
Don't:
Here is the contents of login.ts: [500 lines of code]. Modify it to add rate limiting. Here is register.ts for reference: [300 lines of code].
Exception: If you've already parsed and extracted information (e.g., a specific task from a plan file), provide that directly. Don't make the agent re-read and re-parse a document you've already processed.
Every token in a prompt that the agent doesn't need is waste. Include:
Don't include:
Tell the agent exactly what "done" looks like:
Don't give an agent a grab bag of loosely related work. If you have 3 tasks, spawn 3 agents. Each agent should have a single clear objective.
Use cheaper, faster models for worker agents when the task is straightforward. Reserve the most capable model for orchestration and complex reasoning:
When agents need to work on the same repo without conflicting, use worktrees:
isolation: "worktree" to give an agent its own copy of the repo automaticallygit check-ignore -q <dir>). If it's not ignored, add it to .gitignore first — prevents accidentally committing worktree contents.One attempt, then escalate. The orchestrator should make a single best-effort attempt to resolve an issue. If it's not resolved within a few messages, stop and bring the user in. Burning tokens on retry loops is worse than asking for help.
npx claudepluginhub aeriondyseti/aeriondyseti-plugins --plugin code-flowGuides coordination of parallel agent teams (Implementer, Reviewer, Architect) for complex plans, cross-cutting concerns, and communicating work streams.
Guides composing agent teams for complex tasks with predefined roles: Orchestrator, Explorer, Frontend (React), Backend (Rails), iOS/Swift, Android/Kotlin, Database.
Orchestrates Claude Code agent teams for parallel multi-agent workflows on complex tasks using TeamCreate, SendMessage, TaskUpdate, and shared task lists. Enable with --enable-teams flag.