delegate

Snapshots the full current context — open file, conversation history, project memory files, decisions made, and what's left — then hands the task off to a specialized subagent (preferred) or general-p

SKILL.md

GitHub

delegate

Delegates the current task to a general-purpose subagent with enough context to complete it independently. Unlike a focused specialist, this subagent has full tool access and drives the task to completion — no further prompting needed.

When to Use

  • You want a subagent to own a task end-to-end
  • You need to offload work without losing context
  • A focused skill agent isn't the right tool shape — the task spans files, tools, or domains

Workflow

Step 1 — Detect the Task File

The user has a file open in the IDE. Read it first — it's the anchor for context.

Read the file from ide_opened_file (already provided in the prompt as a system-reminder). Extract:

  • The file path
  • Any relevant content (look for task descriptions, TODOs, comments about what needs doing)

Step 2 — Find Project Memory

Determine the project directory. Common patterns:

  • ~/.claude/agents/{slug}/ — Agent/PD project
  • Current working directory or parent CLAUDE.md

Read the following files if they exist (in order of priority):

memory/next-session.md      ← session-start briefing, carries forward context
memory/decisions.md        ← decisions made so far
memory/heartbeat.md        ← recent status / what's been worked on
memory/state.md            ← machine-readable state snapshot
memory/TODOS.md            ← outstanding tasks
CLAUDE.md                  ← project-level instructions and constraints
PROJECT.md                ← project overview and goals

Step 3 — Synthesize the Briefing

Construct a briefing block that answers:

## Task
[What needs to be done — one sentence, then expanded]

## What's Been Done
[Decisions made, code written, files modified — be specific with paths]

## What's Left
[Outstanding tasks, TODOs, open questions]

## Context & Constraints
- Working file: [path]
- Project: [name/slug]
- Any constraints from CLAUDE.md or decisions.md
- Any relevant file paths the subagent needs to look at

## Verify Before Claiming Done
[The standard this task must meet — e.g., tests pass, builds, user confirms]

Be specific. "The app needs auth" is vague. "Add JWT session middleware to the /api/auth route using the existing AuthService — see src/auth/middleware.ts as a reference" is actionable.

If no project memory files exist, fall back to a brief summary based on the conversation history up to this point. State what's clear and what the subagent should infer.

Step 4 — Spawn the Subagent

Use the Agent tool — check Agency catalog first. If a named specialist fits the task domain, use it. If not, use subagent_type: general-purpose (this skill's purpose is end-to-end delegation, which general-purpose is suited for).

{
  "description": "Complete: [1-line task summary]",
  "name": "[short-slug-for-task]",
  "prompt": "[full briefing block from Step 3]",
  "subagent_type": "[named specialist — check Agency catalog first; general-purpose if no match]"
}

Use run_in_background: true for long-running tasks, false (default) for tasks that return a result to report back.

Step 5 — Report

Report back to the user:

  • What the subagent is doing
  • Its name (for reference if the user wants to message it directly)
  • That it will drive to completion independently

Briefing Quality Checklist

  • Working file path included
  • Task goal is explicit and unambiguous
  • What's been done is specific (file paths, not just "work in progress")
  • What's left is a clear list, not vague
  • Any constraints from CLAUDE.md are called out
  • Verification criteria are stated

Edge Cases

Situation Behavior
No memory files exist Synthesize from conversation history; note uncertainty
ide_opened_file is blank/unclear Infer task from conversation; note assumption
Task is trivial (1-2 steps) Still delegate — use brief briefing, keeps context clean
Subagent spawn fails Report error; suggest running task directly
User interrupts No special handling — subagent continues in background

Notes

  • The subagent does not report back to you automatically — it completes and the user sees the result.
  • If you need to send a message to the running subagent, use SendMessage({to: "[name]"}).
  • For running background tasks, you will be notified when it completes.
  • Do NOT send a shutdown message to the subagent unless the user explicitly requests it.