# Subagent Delegation

> Subagent delegation workflow: proactively move bounded, verifiable side work to explorer/worker/reviewer subagents while keeping the top-level session responsible for integration and verification. Use when a task has independent read-only searches, disjoint implementation slices, review surfaces, or parallel checks; load before spawning or managing multiple subagents.

- Skill: `joernstoehler-dnd-claude-code/subagent-delegation` (Agent Skill)
- Install (CLI): `npx skillmds@latest add joernstoehler-dnd-claude-code/subagent-delegation`
- Raw SKILL.md: https://api.skillmd.com/api/skills/joernstoehler-dnd-claude-code/subagent-delegation/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Integrations & APIs
- Author: JoernStoehler (https://skillmd.com/u/joernstoehler-dnd-claude-code)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/joernstoehler-dnd-claude-code/subagent-delegation

---


# Subagent Delegation

Use subagents for bounded first-pass labor after the active task or focus surface is clear. The top-level session keeps integration and correctness ownership.

Treat delegate output as evidence to check. Before presenting a delegate claim as fact, verify it against files, commands, sources, tests, or a bounded review result.

## Delegate Or Keep Local

Delegate only when the result matters. If the result would not change the main task, skip the subtask.

Then check whether the path matters:

- If only the final result matters, delegate a bounded task with a clear output contract.
- If the path matters for the main thread's reasoning, keep it local or split the work so the main thread sees the needed intermediate state.
- If part of the path matters, make that intermediate state part of the delegate's required output.

Delegate when the subtask is:

- Concrete, self-contained, and easy to state.
- Bounded by a result the main thread can use without seeing every step.
- Verifiable from files, commands, or source citations.
- Bounded by a read-only surface or disjoint write scope.
- Clear enough that the delegate can plan without guessing the task boundary.
- Unblocked, or the blocker-resolution step is explicit in the delegate task.
- Useful because it keeps distracting context, long logs, large source surfaces, or try-and-verify loops out of the main thread.

Keep work local when:

- The path to the result matters and cannot be summarized as required intermediate output.
- The task needs tight context, canon judgment, SotS rules interpretation,
  table-experience judgment, or taste.
- The subtask has no bounded output contract, or the delegate would have to choose the task boundary.
- The result cannot be verified cheaply enough for the top-level session to own it.
- A failed or confused delegate would cost more than doing the work locally.

When the path matters but still contains delegable labor, split the task:

- Have the main thread choose the path, then delegate execution.
- Delegate the first part, require the intermediate state in the output, then continue locally or delegate the second part.
- Redefine the delegate output to include the final result plus the intermediate evidence the main thread needs.

Use narrow subagents proactively for low-risk side work.

Ask Jörn before delegation changes the approved task surface, task ownership,
campaign direction, canon, player-facing framing, or merge readiness.

## Delegate Brief

Each subagent prompt should name:

- Approved surface.
- Objective.
- Required cwd: usually the parent session checkout or assigned worktree; sometimes a subagent-specific worktree.
- Scope: files, directories, or question.
- Ownership: read-only or named write scope.
- Known blockers or blocker-resolution step.
- Success check.
- Output format.
- Decisions reserved for the main thread or Jörn.
- Stop condition.

For repo-edit worker prompts, say it directly. Example:

```text
Required cwd/worktree: <required-cwd>

Use this worktree for all commands and edits. Do not edit `main` or the root checkout.

If you make or detect tracked edits outside <required-cwd>, revert them immediately when that revert is clearly safe. Otherwise stop and report the exact files before continuing.

You are not alone in the codebase; do not revert or overwrite changes made by others.
```

Do not duplicate the same unresolved task across delegates.

For a delegation surface that Jörn may copy into new worktree sessions, use numbered work packets. Keep each packet self-contained enough that the receiving session can start without reconstructing the conversation:

```markdown
1. **Unit of work**
   - Worktree: `short-branch-name`
   - Tracked at: `TASKS.md` entry or "untracked cleanup"
   - Scope: files, directories, or question.
   - Context: why this work matters now and what evidence led to it.
   - Why shallow: what makes this likely agent-doable, or where it may become deep.
   - Decision points: choices reserved for the main thread or Jörn.
   - Expected output: patch, note, report, data file, or finding.
   - Verification: exact command, scan, build, or review check.
   - Stop condition: when to hand back instead of guessing.
```

Use the packet format when proposing several independent tasks, when Jörn is likely to launch sessions manually, or when the top-level session is producing a queue rather than a single subagent prompt. Omit fields only when the answer is obvious from the surrounding packet.

## Running Delegates

- Start sidecar delegates for non-blocking search, review, verification, or disjoint edits.
- While delegates run, continue useful non-overlapping local work.
- Wait only when their result is needed for the next step.
- Verify delegate claims against files, commands, or sources before presenting them as facts.
- For multiple tasks, default to a serial queue: brief A, integrate and verify A, then brief B. Parallelize only independent subtasks.

## Verification Labor

Verification labor can be delegated; verification responsibility cannot.

Good delegated verification tasks:

- Review a file for clear writing, stale paths, rules/source mismatches,
  player/GM-facing leakage, generated-artifact drift, or convention violations.
- Run a command and report the exact failure or success.
- Compare a claim against `TASKS.md`, `library/sots/`, `sessions/`, handoffs,
  scripts, or generated artifacts.
- Inspect a patch for a named risk surface.

The top-level session decides whether the evidence is enough. When reporting verification, name the evidence the top-level session checked and state any residual risk.

## Parallel Delegation

Parallelize for speed only when the subtasks are independent:

- Separate read-only questions.
- Separate review surfaces.
- Disjoint write scopes with named file ownership.
- Independent verification checks that can return in any order.

Do not parallelize shared design decisions, shared write scopes, or dependent tasks. Split by independent surface, then merge results only after local verification.

When running a packetized implementation round, do not spend every live agent
slot on workers. Leave room for at least one reviewer or one bounded cleanup
follow-up.

Name the main thread as integration owner. Parallel delegates report evidence or patches; they do not reconcile conflicts, choose priorities, or change the approved surface.

## Integration

When a delegate returns:

- Check the referenced files or commands before accepting findings.
- Check that the result matches the main task's actual need, not only the delegate prompt.
- Apply or refine worker changes in the main thread if needed.
- Record blockers as concrete missing evidence or Jörn-only decisions.
- Close agents that are no longer needed.