delegation-strategy

An Agent Skill that teaches coding agents when to delegate — and, just as often, when not to.

Agents with subagent tools tend to fail in one of two directions: they spawn agents for lookups they could answer with a single grep, or they grind through forty-file sweeps inline and pollute their own context. This skill gives the agent a four-rung ladder and tells it to pick the lowest rung that fits.

The ladder

Every step up costs latency, tokens, and context-transfer overhead — a subagent starts with zero knowledge of the conversation. The skill spells out the heuristics for each rung: which searches earn a subagent, why single-fact lookups never do, how to write the hand-off prompt, when parallel fan-out pays (and how to keep parallel writers from clobbering each other), and the two hard gates a Workflow must pass (explicit user opt-in plus a task shape that needs deterministic orchestration).

The full text is in skills/delegation-strategy/SKILL.md — about 50 lines, loaded only when the agent is planning a task that could involve delegation. Codex and ChatGPT-compatible skill metadata lives in skills/delegation-strategy/agents/openai.yaml, the Codex plugin manifest lives in .codex-plugin/plugin.json, and the Claude Code plugin manifest lives in .claude-plugin/plugin.json.

Install

Codex / ChatGPT-compatible plugin

This repo is a Codex plugin as well as a standalone Agent Skill. The plugin manifest is at .codex-plugin/plugin.json and exposes skills from ./skills/.

This is a Codex plugin bundle for skills. It is not a legacy ChatGPT ai-plugin.json plugin or a custom GPT Action, because this skill does not expose an external API.

Clone the repo, then install or add it through your Codex plugin flow as a local or Git-backed plugin. The plugin name is mythos-delegation-skill; the bundled skill remains delegation-strategy.

After install, ask Codex to use $delegation-strategy, or let it trigger implicitly when the task involves subagents, parallel review, large searches, or orchestration decisions.

Agent Skills / skills.sh

With the skills.sh CLI (the skill installs under its spec name, delegation-strategy):

npx skills add forjd/mythos-delegation-skill

or with Bun:

bunx skills add forjd/mythos-delegation-skill

Add -g to install user-level (active in every project) rather than project-level.

Manual Codex skill install

For a direct skill install without the plugin wrapper, copy or symlink the skill directory into your Codex skills folder:

git clone https://github.com/forjd/mythos-delegation-skill.git
mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills"
ln -s "$(pwd)/mythos-delegation-skill/skills/delegation-strategy" \
  "${CODEX_HOME:-$HOME/.codex}/skills/delegation-strategy"

Claude Code plugin (marketplace)

This repo is also a Claude Code plugin and its own plugin marketplace (.claude-plugin/plugin.json and .claude-plugin/marketplace.json). Inside Claude Code:

/plugin marketplace add forjd/mythos-delegation-skill
/plugin install mythos-delegation-skill@mythos-delegation

or from the shell:

claude plugin marketplace add forjd/mythos-delegation-skill
claude plugin install mythos-delegation-skill@mythos-delegation

The plugin's skills/ directory is auto-discovered, so installing the plugin registers the delegation-strategy skill.

Manual Claude Code skill install

To install the skill directly without the plugin, clone this repo and symlink the skill directory into Claude Code's skills directory. For user-level install:

git clone https://github.com/forjd/mythos-delegation-skill.git
ln -s "$(pwd)/mythos-delegation-skill/skills/delegation-strategy" ~/.claude/skills/delegation-strategy

For a single project, copy or symlink delegation-strategy/ into that repo's .claude/skills/ instead.

How it works

Agent Skills load progressively. At startup the agent sees only the skill's name and description (~100 tokens); the description is written to trigger when the agent is planning work that could involve searching many files, parallel strands, or large fan-out. Only then does the agent load the full instructions.

The skill is distilled from delegation guidance in Anthropic's Claude Fable 5 harness prompt, then generalized for Agent Skills-compatible runtimes including Claude Code and Codex. It's most useful for agents that don't carry this guidance natively: subagents of subagents, SDK-built agents, and other Agent Skills-compatible tools.

The ladder degrades gracefully. Agents without a workflow tool emulate rung 4 with batches of parallel subagents (the user-opt-in rule still applies); agents with no subagent tools at all get the rung-1 principles — context hygiene, and telling the user when a task is really a fan-out job their tooling can't express.

Repository layout

mythos-delegation-skill/
├── .claude-plugin/
│   ├── plugin.json       # Claude Code plugin manifest
│   └── marketplace.json  # Claude Code plugin marketplace (this repo as a marketplace)
├── .codex-plugin/
│   └── plugin.json       # Codex / ChatGPT-compatible plugin manifest
├── delegation-strategy/
│   └── -> skills/delegation-strategy
├── skills/
│   └── delegation-strategy/
│       ├── agents/
│       │   └── openai.yaml   # Codex / ChatGPT-compatible skill metadata
│       └── SKILL.md          # the skill: frontmatter + instructions
├── LICENSE
└── README.md

Development

The skill follows the Agent Skills specification. To validate after editing:

npx skills-ref validate ./skills/delegation-strategy

To validate the Claude Code plugin and marketplace manifests:

claude plugin validate .

For Codex compatibility, also run the Codex skill validator if available:

python3 /path/to/skill-creator/scripts/quick_validate.py ./skills/delegation-strategy
python3 /path/to/plugin-creator/scripts/validate_plugin.py .

Codex uses only the name and description frontmatter fields to decide when to load a skill, so keep trigger wording in description, UI-facing skill metadata in skills/delegation-strategy/agents/openai.yaml, and plugin-facing metadata in .codex-plugin/plugin.json.

Issues and PRs welcome — especially counterexamples where the ladder gives the wrong answer.

License

MIT