Plan → Review → Act

Problem

Developers dump requirements into the AI prompt and ask it to implement them. Sometimes this works out of the box and there are no problems. Sometimes it doesn’t.

The three-phase loop

┌─────────┐     ┌──────────┐     ┌─────────┐
│  PLAN   │  →  │  REVIEW  │  →  │  ACT    │
│         │     │          │     │         │
│ AI maps │     │ You check│     │ AI      │
│ the work│     │ the plan │     │ executes│
└─────────┘     └──────────┘     └─────────┘
                      │
                 ← Adjust plan

Also known as Refine-Plan-Act. It is the same workflow, same idea but a different name.

Phase	AI does	You do
Refine (pre-check)	• Clarifies requirements against codebase • Detects missing or inconsistent assumptions	• Answer questions • Confirm corrected requirements
Plan	• Identifies files to change • Produces step-by-step implementation plan	• Review plan critically • Check for wrong assumptions
Review	• No action	• Decide whether to revise plan or proceed
Act	• Edits files • Runs commands • Creates PRs	• Monitor execution • Interrupt if behavior deviates

Tip: Run each phase in a fresh session. Save the plan as a file (e.g. PLAN.md) — it keeps thinking separate from code, so if "Act" goes wrong you restart from a known-good plan, not from scratch.

Why Review comes before Act

The loop is Plan → Review → Act, not Plan → Act → Review.

Review after acting:

• Mistakes are discovered only after damage is done
• You may need to revert multiple files or changes
• Side effects may already have happened (git commits, API calls, data changes)
• Errors can cascade into further incorrect actions
• Debugging becomes reactive instead of preventive
• Trust in the process decreases after each mistake

Review before acting:

• Mistakes are caught before anything is changed
• Only a single step or sentence needs adjustment
• No side effects occur yet (safe checkpoint)
• Prevents incorrect commits, requests, or data modifications
• Keeps the system in a controlled, predictable state
• Builds trust step by step through safe execution

Signals that the Plan phase needs more work

Signal	Meaning	Action
Plan changes too many files	Scope too broad	Reduce scope, split into smaller tasks
Mentions non-existent files	Hallucinating codebase	Add context (file tree / relevant files)
Skips test step	Risk of silent failures	Require tests after each change
Vague instructions	Not actionable	Ask for exact methods and reasons

References

• The Refine-Plan-Act Pattern for Agentic AI Coding — Francesco Borzì