Dialog Engineering

Dialog Engineering

Single prompts fail at complex problems. Structured dialog succeeds.

The CSAR Loop

Every effective AI conversation follows four phases in a loop:

  Clarify ──→ Summarize ──→ Act ──→ Reflect
     ↑                                  │
     └──────────────────────────────────┘

The loop is not one-shot. Complex tasks cycle 3-5 times. Each cycle narrows the solution space.

Why Dialog Beats Single Prompts

Turn Design Patterns

Turn 1: The Anchor (Clarify + Summarize)

State context AND goal in the first turn. Include:

• Tech stack and constraints
• What output you want (code, doc, plan, analysis)
• What you DON'T want (avoids the most common waste)

Good: "I need to add authentication to my Express/TypeScript API. Current stack: Express 4, Prisma, PostgreSQL. What approaches do you recommend? Keep it simple."

Bad: "Build me a complete user authentication system with login, registration, password reset, JWT tokens, refresh tokens, email verification, rate limiting, and tests."

Turn 2+: Steering Moves

The Closing Turn (Reflect)

End sessions by asking:

• "What did we miss?"
• "What would break this?"
• "Summarize the decisions we made"

This catches blind spots and creates a record for future sessions.

Dialog Anti-Patterns

When to Use Each CSAR Phase

CSAR Implementation

// Implement CSAR loop tracking in conversation
enum CSARPhase {
  Clarify = 'clarify',
  Summarize = 'summarize', 
  Act = 'act',
  Reflect = 'reflect'
}

interface DialogState {
  currentPhase: CSARPhase;
  cycleCount: number;
  contextGathered: Map<string, string>;
  goalStatement: string | null;
  lastAction: string | null;
}

function detectPhaseFromMessage(message: string): CSARPhase {
  // Clarify signals: providing context
  if (message.match(/I'm building|current stack|constraint is|using/i)) {
    return CSARPhase.Clarify;
  }
  // Summarize signals: stating goal
  if (message.match(/show me|I need|the goal is|help me/i)) {
    return CSARPhase.Summarize;
  }
  // Reflect signals: evaluation/iteration
  if (message.match(/what did|go deeper|what if|why did you/i)) {
    return CSARPhase.Reflect;
  }
  // Default: Act phase (generation expected)
  return CSARPhase.Act;
}

function suggestNextMove(state: DialogState): string {
  switch (state.currentPhase) {
    case CSARPhase.Clarify:
      return state.goalStatement 
        ? 'Ready to act on your goal.'
        : 'What specifically do you want to accomplish?';
    case CSARPhase.Act:
      return 'Shall I go deeper on any aspect?';
    case CSARPhase.Reflect:
      return state.cycleCount < 3 
        ? 'Want to refine further, or ready to proceed?'
        : 'We've iterated well. Ready to finalize?';
    default:
      return '';
  }
}

Integration with the AI assistant Skills

• Meditation: The Reflect phase maps to meditation's "what did I learn?" step
• Research-first-development: Clarify phases should invoke research skills before acting
• Code review: Reflect phase should invoke adversarial review checklist
• Knowledge synthesis: End-of-session Reflect should capture insights for global knowledge