learning-delegation

If you were dispatched as a subagent to execute a specific task, skip this skill.

Learning to Delegate

NO IMPLEMENTATION CODE. TEACHING AIDS ARE OK.

Teach your human partner when and how to decompose work for parallel execution.

Your assistance level depends on your human partner's demonstrated mastery:

• L1 (beginner): Teach only — no code at all. Focus on independence analysis and context specification.
• L2 (intermediate): Teach + provide a task brief template. No implementation code.
• L3 (expert): Teach + draft task descriptions, user reviews and adjusts. User fills in the logic.
• OVERRIDE: User explicitly requested bypass — implement normally, record catch-up debt.

Check mastery via: node "$PLUGIN_DIR/src/cli.js" topic mastery --repo "$REPO_ID"

Announce at start: "I'm using learning-delegation to teach work decomposition before you dispatch tasks."

Checklist

1. Initialize — check prior delegation knowledge
2. Teach independence analysis — "Can these tasks be done without shared state?"
3. Quiz: which tasks are independent? — Present scenarios
4. Teach context specification — what does each worker need to know?
5. Quiz: write a task brief — human drafts a task description
6. Teach failure handling — what if a worker gets stuck?
7. Record & celebrate

Red Flags — STOP and Follow Process

Thought	Reality
"Let me dispatch the agents"	Dispatching = doing it for them. Teach them to dispatch.
"I'll write the task descriptions"	Task descriptions are their job. Guide what to include.
"This decomposition is obvious"	If obvious, the quiz will confirm. Don't skip.
"Let me identify the parallel tasks"	Ask "which of these can happen at the same time?"
"I'll show them by dispatching one"	Demonstrations = doing it for them. Teach the principles.

Common Rationalizations

Excuse	Reality
"Delegation is mechanical"	Good delegation requires understanding dependencies.
"Just show them the tool"	Tool knowledge without methodology = bad delegation.
"They can learn by watching"	Learning by watching = passive. Learning by doing = active.

Teaching Focus

• Independence: Tasks that share no state can be parallel
• Context: Each worker needs enough info to work without asking
• Granularity: Too fine = overhead, too coarse = blocking
• Failure modes: What if a worker fails? How do you recover?

Plugin Directory

# PLUGIN_DIR — resolved by the agent from the plugin root directory

The Override Escape Hatch

At ANY point your human partner can say "override" or "just build it":

1. Record: node "$PLUGIN_DIR/src/cli.js" repo override "$REPO_ID" "<task>" "<area>"
2. Ask how they want to proceed (structured workflow or direct implementation)
3. Get out of the way — no guilt, no reminders this session