English · Português · Español · 简体中文

delegate

A Claude Code skill that turns a messy prompt into a clean brief, picks the right depth tier and the cheapest agent that can do the job, dispatches it, and shows you what it cost.

The point is not "spawn a subagent". Claude Code already does that. The point is that naive delegation is expensive — a subagent re-sends its entire system prompt and its whole tool-schema block on every turn — and this skill makes the cheap path the default path.

you:      /delegate the login redirect is broken after the session refactor

claude:   agent:  delegate-deep
          model:  opus
          effort: deep — "Think hard. Trace the full flow before editing..."
          prompt:
          Fix the post-login redirect regression introduced by the session refactor.
          Start at src/auth/session.ts and src/routes/login.tsx. Follow AGENTS.md.
          You have no shell — put validation commands into `commands_to_run`.

          ...

          route:  A (claude -p)
          agent:  delegate-deep · tools Read,Edit,Write,Grep,Glob
          model:  opus · effort high
          tokens: 14 in · 17.2k cache write · 24.2k cache read · 65 out · 7 turns
          cost:   $0.42 of $2.00 budget · 1m 12s

What it actually does

Refuses cheap work	Step 0 asks whether delegating is worth it at all. A 10-line edit costs more delegated than done inline — the skill says so and just does it.
Rewrites the prompt	The subagent has zero conversation context. Your prompt becomes a short standalone brief with the file paths already named, because a vague brief costs far more in a subagent than in your session.
Routes by depth, not by model	Three tiers — trivial / standard / deep. The model stays pinned; what changes is reasoning effort.
Picks a narrow agent	general-purpose, Explore and Plan all carry every MCP server you have connected, re-sent every turn. The bundled delegate-scout / delegate-worker / delegate-deep carry none.
Strips MCP entirely on route A	claude -p --strict-mcp-config loads zero MCP servers: measured 17.1k context per turn vs 44.0k with the default config.
Withholds the shell	A subprocess can't surface a permission prompt, so the subagent gets no Bash. It returns the commands it wants run and your session runs them, where permissions still work.
Prints the bill	Every run closes with a card: tokens, turns, cost against budget, wall time. So you can see which route is actually cheaper instead of guessing.

Install

Requires Claude Code.

git clone https://github.com/ocauapaz/claude-delegate.git
cd claude-delegate
./install.sh

./install.sh --project installs into ./.claude of the current repo instead of ~/.claude, if you only want it in one project.

Windows PowerShell (no bash)

git clone https://github.com/ocauapaz/claude-delegate.git
cd claude-delegate
New-Item -ItemType Directory -Force "$HOME\.claude\skills","$HOME\.claude\agents"
Copy-Item -Recurse -Force .\skills\delegate "$HOME\.claude\skills\"
Copy-Item -Force .\agents\delegate-*.md "$HOME\.claude\agents\"

Manual — no git

Download the repo as a zip, then copy:

• skills/delegate/ → ~/.claude/skills/delegate/
• agents/delegate-*.md → ~/.claude/agents/

Restart Claude Code afterwards so it picks up the new skill and agents.

Usage

/delegate <whatever you were about to type>

It also fires on plain phrasing — "hand this to a subagent", "get an agent to do this".

Nothing else to learn. The skill decides the tier, the agent and the route; you read the dispatch block before it runs and the run card after.

If a run comes back thin, say so — the skill re-dispatches one tier up with the failure output pasted into the brief.

Configuration

Two knobs, both inside skills/delegate/SKILL.md:

• Default model (step 2 table) — ships as opus, with haiku for the trivial tier. Change either if your plan or budget differs; the rest of the skill doesn't care which model it names.
• --max-budget-usd (step 4) — ships as 2, a hard ceiling on route A. Raise it for deep work, lower it if you want a tighter leash.

The three agent definitions in agents/ are also plain markdown — edit their effort, tools or instructions to taste.

Routes

Route A — claude -p. Preferred. A real --effort knob, zero MCP schemas, a hard dollar ceiling, and structured JSON back including full usage data. Requires claude auth status to report "loggedIn": true; if not, run claude setup-token once.

Route B — the Agent tool. Fallback. No effort parameter (the skill injects a depth line into the prompt instead) and no usage data, so the run card reports cost: n/a (route B) rather than guessing.

Layout

skills/delegate/SKILL.md    the skill
agents/delegate-scout.md    read-only investigator     (Read, Grep, Glob, Bash)
agents/delegate-worker.md   scoped implementer         (+ Edit, Write)
agents/delegate-deep.md     high-effort implementer    (+ Edit, Write, more thinking)
install.sh

License

MIT