ai-prompting-patterns

AI Prompting Patterns & Agent Communication

Core Principle

An agent that acts confidently on wrong assumptions costs more than an agent that pauses to ask.

The goal is not maximum autonomy — it's maximum useful autonomy. An agent should be autonomous enough to not constantly interrupt, but disciplined enough to catch itself before doing something irreversible or wrong.

---

The Autonomy Dial

Set the right autonomy level based on task risk:

Level	Mode	When to Use
1	Observe & Suggest	High-risk environment, new agent, first interactions
2	Plan & Propose	Complex tasks — agent creates plan, every step needs approval
3	Act with Confirmation	Agent prepares full action sequence, asks for final go/no-go
4	Act Autonomously	Pre-approved, low-risk, reversible tasks only

Default: Level 2 for complex tasks, Level 3 for well-defined tasks, Level 4 only for formatting/style changes.

---

Intent Clarification

When to Ask (Ask-when-Needed Protocol)

Ask before acting when:

• A key parameter is missing
• There are multiple valid interpretations
• The task spans multiple files or systems
• The action could be hard to reverse
• Confidence in understanding is below threshold

Don't ask when:

• The task is simple and single-responsibility
• Context makes intent obvious
• The question can be answered by reading the codebase

How to Ask

Use the Binary/Choice pattern — propose specific options instead of open-ended questions:

❌ Bad: "What do you want me to do?"
✅ Good: "I can implement auth using either JWT or OAuth2. 
          Given oauthlib is already in your requirements.txt, 
          I recommend OAuth2. Should I proceed with that?"

Clarification Template

Before I proceed, I want to confirm my understanding:
- I interpret this task as: [your interpretation]
- Key assumption I'm making: [assumption]
- If this is correct, I'll [action]. If not, please clarify.