Do In Steps

Execute a complex task as a chain of focused sub-agents, each building on the previous step's output.

Steps

1. Decompose the Task

Break the request into discrete, ordered steps. Each step must have:

• A single clear objective
• Explicit inputs (what it receives from the prior step)
• Explicit outputs (what it passes to the next step)
• A model assignment based on complexity

Model Routing

Complexity	Model	Use For
Simple	haiku	Lookups, reads, searches, summarization
Standard	sonnet	Implementation, file edits, refactoring
Complex	opus	Architecture, deep analysis, ambiguous design

2. Execute Each Step Sequentially

For each step N:

1. Announce: [Step N/Total] <Objective> — dispatching <agent> (<model>)
2. Dispatch a Task agent with:
   • The step objective
   • A concise summary of all prior step outputs (not raw dumps)
   • Relevant file paths, symbol names, and code context
   • Explicit scope: what this agent should and should not do
3. Wait for the result before proceeding
4. Verify the result meets the step's output criteria
5. If verification fails: diagnose, re-dispatch once with a corrected prompt, then escalate on second failure

Never dispatch step N+1 until step N is verified.

3. Pass Context Forward

Summarize each completed step into a compact handoff:

Step N result: <2-3 sentence summary of what was done and key outputs>
Artifacts: <file paths or symbols changed/created>

Include this handoff block in every subsequent agent prompt. Keep it under 10 lines total regardless of how many steps have completed.

4. Synthesize

After all steps complete, deliver a unified summary:

• What each step accomplished
• All files or symbols changed
• Any deferred issues or open questions

Rules

1. Each agent owns a distinct scope — no two parallel agents touch the same file
2. Always include file paths and symbol names in agent prompts — never spawn without context
3. After 2 failed verifications on the same step, escalate to oh-my-claudecode:architect
4. Do not compress the handoff so aggressively that the next agent lacks needed context

Output Format

## Step Plan
Step 1: <Objective> [haiku/sonnet/opus]
Step 2: <Objective> [haiku/sonnet/opus]
...

## Execution
[Step 1/N] <Objective> — complete
  Result: <summary>
[Step 2/N] <Objective> — complete
  Result: <summary>
...

## Summary
<What was accomplished end-to-end>

## Open Issues
<Unresolved items, or "None">