A focused agent skill for executing approved plans with scoped subagents, clear file ownership, concurrency limits, branch/worktree discipline, progress tracking, and review gates.
This repo packages one reusable skill:
skills/subagent-driven-development/: the Codex skill for multi-agent implementation work.assets/title-cover.png: GitHub/social preview image for this repo.
Large implementation plans break down when agents overlap, drift from the plan, write into the same checkout, or keep piling context into one long thread.
subagent-driven-development gives your coding agent a practical operating model:
- fresh subagent per bounded task
- 2-3 active code-edit workers per real integration lane
- one writer per physical checkout/worktree
- protected dirty files
- human-readable branch names only
- visible ledger and terminal progress bars
- review gates sized to the risk of the task
- controller-owned integration and final review
The easiest install path is to give this repo to your coding agent and let it copy the skill into the right local folder.
Copy this prompt into Codex, Claude Code, or another local coding agent:
Install the Subagent-Driven Development skill into my local agent setup.
Repo: https://github.com/shane9coy/subagent-driven-development
Tasks:
1. Clone or download the repo.
2. Copy this skill folder into my agent skills folder:
- skills/subagent-driven-development
3. Default Codex skills target: ~/.codex/skills/
4. Claude Code skills example: ~/.claude/skills/
5. Do not overwrite an existing subagent-driven-development skill without showing me what will change first.
6. Verify the installed SKILL.md and agents/openai.yaml files exist and report the final paths.
This repo intentionally does not ship an AGENTS.md file. It works best beside the canonical AGENTS.md from the Agent Skills Architecture Guide, because this skill was shaped around those same operating rules: clean branches, scoped work, progress updates, verification, and review discipline.
Start there if you want the full base setup, including:
AGENTS.md: global agent behavior, Git hygiene, and handoff rulesagent-copy: copy canonical instructions into project reposnew-skill-builder: create, install, and validate skillsnew-mcp-builder: plan and build MCP integrations
Then add this skill when you want optimized subagent planning and execution.
Codex is the default example. Run from the directory where you want to download the repo:
git clone https://github.com/shane9coy/subagent-driven-development.git
cd subagent-driven-development
mkdir -p "$HOME/.codex/skills"
cp -R skills/subagent-driven-development "$HOME/.codex/skills/"Verify:
test -f "$HOME/.codex/skills/subagent-driven-development/SKILL.md"
test -f "$HOME/.codex/skills/subagent-driven-development/agents/openai.yaml"Claude Code users can copy the same folder to their configured skills directory, commonly:
~/.claude/skills/subagent-driven-development
After install, the skill is easy to call from Codex. Type $sub in the prompt box and Codex should surface subagent-driven-development as an available skill.
It works well in normal prompts, but it is most useful while Codex is building or refining a plan. Calling it during planning helps Codex break the work into cleaner segments before execution, and gives you a clearer view of what each agent will own.
Use $subagent-driven-development while refining this plan so the implementation can be split into scoped subagent tasks.
That tells Codex to think in terms of controller work, agent ownership, progress tracking, and review gates before the implementation starts.
- Start with an approved implementation or remediation plan.
- Use this skill only when the work has independent tasks worth delegating.
- Let the controller inspect git state, protect dirty files, define ownership, and assign agents.
- Keep shared contracts, lockfiles, migrations, generated code, and global config serial.
- Integrate subagent outputs in dependency order.
- Run the right review gate: lightweight combined review for trivial edits, separate spec and quality reviews for non-trivial work.
subagent-driven-development/
|-- LICENSE
|-- README.md
|-- assets/
| |-- title-cover.png
| `-- subagent-driven-development-repo-preview.png
`-- skills/
`-- subagent-driven-development/
|-- SKILL.md
`-- agents/
`-- openai.yaml
Use subagent-driven-development when executing an approved plan with independent tasks.
It guides the controller to:
- extract tasks and order them by dependency
- assign bounded scopes and reasoning effort
- maintain an agent ledger
- keep branch/worktree ownership clean
- use deterministic progress stages
- dispatch fresh fix agents when needed
- run final integration review before handoff
It guides subagents to:
- work only within assigned ownership
- avoid protected user changes
- list files changed
- run required verification
- return a compact status, risk, and follow-up report
| Tool | Suggested skill folder |
|---|---|
| OpenAI Codex | ~/.codex/skills/subagent-driven-development |
| Claude Code | ~/.claude/skills/subagent-driven-development |
| AgentSkills-compatible tools | configured skills directory |
| Repo-local use | skills/subagent-driven-development |
Created by @shaneswrld_.
MIT - use it, fork it, ship it. Credit is appreciated.