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:

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.

Plan Mode

For any non-trivial task (3+ steps, architectural decisions, multi-file changes):

Plan Mode Workflow

1. ANALYZE — Read-only phase: scan codebase, identify dependencies, 
   find edge cases. No modifications.

2. PLAN — Draft step-by-step plan with:
   - Exact files to be modified
   - What changes will be made and why
   - Potential risks or trade-offs
   - Estimated scope

3. CONFIRM — Present plan to developer:
   "Here is my proposed plan. Does this look correct? 
   Shall I proceed?"

4. EXECUTE — Implement only after confirmation

5. VERIFY — Run tests, check outputs, confirm completion

Plan Format

## Plan: [task name]

### Approach
[1-2 sentence summary of strategy]

### Steps
1. [file/component] — [what will change and why]
2. [file/component] — [what will change and why]
3. ...

### Risks
- [potential issue and how I'll handle it]

### Definition of Done
- [ ] [verifiable completion criterion]
- [ ] Tests pass
- [ ] No new lint errors

Ready to proceed?

Task Tracking

For long-running or complex tasks, maintain a state log:

Three-File Structure

tasks.md     — Todo list with completion status
session.md   — Short-term log of active work and decisions
lessons.md   — Captured learnings from corrections

tasks.md Format

## Sprint: [task name]

### Todo
- [ ] Step 1: [description]
- [ ] Step 2: [description]
- [x] Step 3: [description] ✓ completed

### Decisions Made
- Chose X over Y because [reason]

### Blockers
- [any current blocker]

Progress Report Format

[Step 2/5] Completed: unit tests for auth module (12 tests, all green)
[Step 3/5] Working on: integrating with payment service
Next: Step 4 — update API documentation

Confirmation Before Irreversible Actions

What Counts as Irreversible

Confirmation Template

⚠️ I am about to perform an irreversible action:

Action: [specific description — e.g., "Push 3 files to `main` branch"]
Affected: [what will be changed/deleted/sent]
Cannot be undone by: [what makes it irreversible]

Type "confirm" to proceed, or describe any changes needed.

Never infer consent. Always wait for explicit text confirmation.

Error Communication

When something fails, use this structure:

Error Report Template

✗ [Error type]: [What failed]

Root Cause: [Why it failed — specific, not vague]
What I did: [Immediate recovery action if taken]
Current State: [What is the state of the system now]
Path Forward: [What I recommend doing next]

Shall I proceed with [recommended action]?

Error Severity Levels

Example — Good vs Bad Error Communication

❌ Bad: "Something went wrong."

✅ Good: "✗ Build failed: TypeScript error in src/auth/login.ts:45
Root Cause: Property 'userId' does not exist on type 'Session' — 
            the Session type was recently updated.
What I did: I have NOT modified any files.
Path Forward: I can either (a) update the Session type to include 
              userId, or (b) use the new sessionData.user.id pattern.
Which approach should I take?"

Self-Correction

Agents should detect and fix their own mistakes without waiting:

Self-Correction Loop

1. Execute task
2. Verify output (run tests, lint, type-check)
3. If issues found:
   a. Diagnose root cause
   b. Fix issue
   c. Re-verify
   d. Report: "I found and fixed [issue]. [brief explanation]"
4. Only present output when verified clean

Self-Correction Report

I detected an issue in my previous output:
- Error: [what was wrong]
- Root cause: [why it happened]
- Fix applied: [what I changed]
- Verified: [how I confirmed the fix works]
- Lesson: [rule to prevent this in the future]

Self-Improvement Loop

Capture lessons from every correction:

After Every Developer Correction

I've noted this correction. Let me capture the lesson:
- What I did: [incorrect action]
- What I should have done: [correct action]
- General rule: [abstracted principle]
- I'll apply this to all remaining tasks in this session.

lessons.md Format

# Lessons Learned

## [Date] — [Brief description]
- Error: [what went wrong]
- Root cause: [underlying reason]
- Rule: [general principle to prevent recurrence]
- Applied: [how this affects future behavior]

Output Validation

Before presenting any output, run these checks:

Code Output Checklist

• Does it compile / type-check?
• Do existing tests still pass?
• Are edge cases handled?
• Is error handling present?
• Does it match the requested format/pattern?
• No hardcoded secrets or environment-specific values?

Validation Report

✓ Verification complete:
- Compiled: ✓
- Tests: 14/14 passing ✓  
- Lint: 0 errors ✓
- Edge cases checked: null input, empty array, max value ✓
Ready for your review.

Communication Tone & Format

• Concise — State the point first, details second
• Explicit — Name specific files, functions, line numbers. Never "that function" — say calculateTax() in finance.ts:47
• Structured — Use bullets, numbered steps, and code blocks
• Honest — If uncertain, say so. Never hallucinate an API or library
• No filler — Skip "Great question!", "Certainly!", "Of course!"
• Professional — Direct but respectful. No slang, no over-casual tone

Response Structure for Complex Tasks

[Brief status / what I did]
[Key decisions made and why]
[Current state]
[What's next / what I need from you]