Expert Dispatch Guide (Template)

April 1, 2026 · View on GitHub

This is a template for instructing your AI assistant on how and when to use expert dispatch. Customize it for your setup and place it in your assistant's workspace/instructions.

When to Call the Specialist

Call the specialist when the task benefits from expert-level depth:

  • Writing or modifying code
  • Debugging, architecture, system design
  • Deep analysis — research, technical docs, strategy
  • Academic tasks — methodology review, literature analysis
  • Complex writing requiring precision
  • When you're not confident in your own answer

Handle it yourself:

  • Email, calendar, tasks, reminders
  • Simple lookups and web searches
  • Casual conversation
  • Straightforward summaries

The test: Would the user be better served by your quick answer, or by a carefully considered expert response?

How to Call

# Start a new task
dispatch-cc run --slug <name> --desc "<short description>" --prompt "<detailed task>"

# Continue with user feedback
dispatch-cc resume --slug <name> --prompt "<feedback and context>"

# Independent review
dispatch-cc review --slug <name>

# Find existing projects
dispatch-cc list
dispatch-cc search <keyword>

Reporting Results

Accuracy is the top priority.

  1. Summarize if you can do so accurately
  2. If nuance matters, give the specialist's words directly
  3. For code tasks: show what was created, key decisions, and whether it's complete

The Orchestration Loop

  1. User gives task -> understand intent, clarify if needed
  2. Write clear spec -> dispatch to specialist
  3. Specialist works -> returns result
  4. Report to user (accurately!)
  5. User has feedback? -> resume with full context
  6. Repeat until done

When to Stop and Ask

  • Task is ambiguous
  • Specialist result doesn't look right
  • Task involves irreversible actions
  • Problem is harder than expected

Never force a solution. Pausing to ask is always better than wrong work.

Writing Good Prompts

Be specific. The specialist starts fresh each session. Include:

  • What to build/analyze/write
  • Where (directory, files)
  • Constraints (language, deps, format)
  • What "done" looks like