Back to all articles
AIWorkflowDevelopmentAgents

Plan, Execute, Critique: The Three Modes of AI-Assisted Development

Stop treating AI as autocomplete. The real power is in modes: planning before coding, executing autonomously, and multi-model critique. Here's how it works.

Plan, Execute, Critique: The Three Modes of AI-Assisted Development

Most developers use AI wrong.

They type a prompt, get code, paste it, fix the bugs, repeat. It's autocomplete with extra steps.

But AI can do so much more when you give it modes.

  • Plan Mode: Think before you code
  • Execute Mode: Autonomous implementation
  • Critique Mode: Multi-model review

This is how AI-assisted development was meant to work.


The Problem with Promptless AI

The Current Workflow

Developer: "Write a function to validate emails"
AI: [generates code]
Developer: [pastes code]
Developer: [tests it]
Developer: "It doesn't handle edge case X"
AI: [generates new code]
Developer: [pastes again]
...repeat 5 times

You're doing the thinking. AI is just typing.

What's Missing

  • No planning — AI doesn't know your codebase
  • No context — Each prompt starts fresh
  • No verification — You're the only reviewer
  • No autonomy — AI waits for every instruction

It's a fancy typewriter, not a pair programmer.


Mode 1: Plan

Before writing a single line of code, the AI maps out the approach.

How Plan Mode Works

You: "Add user authentication to the app"

AI (Plan Mode):
1. Analyzing existing codebase...
2. Found: No current auth, using Next.js, Prisma ORM
3. Recommended approach:
   - Add NextAuth.js for authentication
   - Create User model in Prisma schema
   - Add login/signup pages
   - Protect routes with middleware
   - Store sessions in database
4. Files to create:
   - app/api/auth/[...nextauth]/route.ts
   - app/(auth)/login/page.tsx
   - app/(auth)/signup/page.tsx
   - middleware.ts
5. Files to modify:
   - prisma/schema.prisma
   - lib/auth.ts (new)
6. Estimated changes: 8 files, ~400 lines

Approve this plan?

Why Planning Matters

  • Catches bad approaches early — Before you write 500 lines in the wrong direction
  • Shows the full scope — Know what you're getting into
  • Maintains consistency — AI follows your existing patterns
  • You stay in control — Approve before execution

When to Use Plan Mode

  • New features
  • Major refactoring
  • Unfamiliar territory
  • Anything touching > 3 files

Mode 2: Execute

Once the plan is approved, AI works autonomously.

How Execute Mode Works

Plan approved. Executing...

[Creating] app/api/auth/[...nextauth]/route.ts
[Creating] lib/auth.ts
[Modifying] prisma/schema.prisma
[Running] npx prisma generate
[Creating] app/(auth)/login/page.tsx
[Creating] app/(auth)/signup/page.tsx
[Creating] middleware.ts
[Running] npm run build
[Testing] Auth flow...

✓ Build successful
✓ Login page renders
✓ Signup creates user
✓ Protected routes redirect
✗ Session not persisting (fixing...)
[Modifying] lib/auth.ts - added session callback
[Running] npm run build
✓ All tests passing

Execution complete. 8 files changed.

What Makes Execute Mode Different

  • Multi-step operations — Not just generating code, but running builds, tests, migrations
  • Self-correction — Errors trigger fixes, not just error messages
  • Real integration — Code goes into your actual files, not a chat window
  • Verification — AI confirms the feature works before reporting done

When to Use Execute Mode

  • After plan approval
  • Bug fixes with clear scope
  • Repetitive tasks (generate 20 components)
  • When you trust the plan

Mode 3: Critique

Multiple AI models review the work, debating until consensus.

How Critique Mode Works

Running critique on: user authentication implementation

Agent 1 (GPT-5 - Security Focus):
- ✓ Password hashing looks correct
- ⚠ Rate limiting not implemented on login
- ✗ Session token in URL parameter is vulnerable

Agent 2 (Claude - Code Quality):
- ✓ Clean code structure
- ⚠ Missing input validation on signup
- ⚠ Error messages could leak user existence

Agent 3 (Gemini - Best Practices):
- ✓ Following NextAuth conventions
- ⚠ Should add email verification flow
- ⚠ Missing CSRF protection on forms

Consensus (3/3 agree):
- Must fix: Session token vulnerability
- Should fix: Rate limiting, input validation
- Consider: Email verification, CSRF

Auto-fixing critical issues...

Why Multi-Model Critique Works

  • Different perspectives — Each model has different training, catches different issues
  • Reduced blind spots — What one misses, another catches
  • Consensus signals importance — If all three flag it, it matters
  • Automated fixing — Critical issues get fixed, not just reported

When to Use Critique Mode

  • Before shipping to production
  • Security-sensitive code
  • Complex business logic
  • Code you didn't write yourself
  • After major refactoring

The Three Modes Together

Example: Adding a Payment System

Phase 1: Plan

You: "Add Stripe payment integration for subscriptions"

AI plans:
- Stripe SDK setup
- Webhook handlers
- Subscription model
- Checkout flow
- Customer portal
- 12 files, ~800 lines

You: "Approved"

Phase 2: Execute

AI executes:
- Creates all files
- Sets up Stripe
- Runs test transactions
- Verifies webhooks
- Self-corrects 3 issues
- Reports complete

Phase 3: Critique

AI critique:
- Agent 1: "Webhook signature verification is correct"
- Agent 2: "Missing idempotency keys on charges"
- Agent 3: "PCI compliance looks good"
- Consensus: Add idempotency keys
- Auto-fixed

Result: Production-ready payment system. You wrote zero lines of code.


Implementing the Modes

Without Proper Tooling (Hard Mode)

1. Plan: Manually prompt ChatGPT to analyze your codebase
   (requires pasting code, losing context)

2. Execute: Copy code, paste into files, run commands manually
   (back to being a code monkey)

3. Critique: Open Claude tab, paste code, ask for review
   Open GPT tab, paste same code, ask for review
   Manually compare responses
   (tedious and incomplete)

With a UDE (Easy Mode)

1. Plan: Agent reads your codebase, proposes plan
2. Execute: One click, agent implements everything
3. Critique: Multiple models review automatically

The modes exist. The question is whether your tools support them.


Getting Started

If You're Using ChatGPT/Claude Directly

You can approximate the modes:

Plan prompt: "Before writing any code, analyze my requirements and propose a detailed implementation plan. List all files to create/modify and your approach for each."

Critique prompt: "Review this code for security issues, code quality, and best practices. Be critical. I want to find problems."

But you'll still be copy-pasting and managing context manually.

If You're Using an AI-Native Tool

Look for:

  • Plan mode built in
  • Agent execution (not just generation)
  • Multi-model support for critique
  • Codebase awareness (not just current file)

The Orbit Approach

Orbit implements all three modes natively:

  • Plan mode in Agent tab
  • Execute mode with autonomous operation
  • Critique mode with multi-model consensus

No prompting tricks required.


The Mindset Shift

Stop thinking of AI as:

  • Autocomplete
  • Code generator
  • Stack Overflow replacement

Start thinking of AI as:

  • Architect (Plan mode)
  • Implementer (Execute mode)
  • Reviewer (Critique mode)

You're the technical lead. AI is your team.

Direct, don't type. Review, don't write. Ship, don't struggle.


The Future of Development

The three modes aren't a feature. They're a paradigm.

As AI gets better, the modes get more powerful:

  • Better plans with deeper codebase understanding
  • Better execution with fewer self-corrections needed
  • Better critique with more specialized reviewers

The developers who master the modes now will lead the teams of tomorrow.


Plan. Execute. Critique.

This is how AI-assisted development works.

Everything else is just autocomplete.