Plan-First & Steering
The single biggest lever you have with a coding agent is making it plan before it acts — and then staying in the loop to steer it. You are the driver, not a bystander.
Plan Before Code
Many agents can propose a plan before touching anything. Use it. Catching a wrong approach in a paragraph is far cheaper than catching it in code you now have to unpick.
- Ask for a plan — a numbered list of the steps it intends to take.
- Read it — does it match what you actually want?
- Correct it — fix the approach while it is still just words.
- Then execute — let it build only once the plan is right.
Steer Mid-Task
A plan is not a contract you walk away from. As the agent works you can interrupt, redirect, and add constraints — the moment you see it drift, not after.
Small Steps, Checkpoints
An agent left to run unchecked tends to wander. A few simple habits keep it honest.
- Small steps — one focused change at a time beats one giant leap.
- Checkpoints — pause and review before moving on.
- “Show me first” — ask it to explain a big change before making it.
The Same Loop
This is nothing new — it is the build → inspect → decide loop from Module 1, applied to an agent. You build a plan, inspect what it produces, and decide whether to keep going or steer.
Build It
How to implement: on your next task, ask the agent for a step-by-step plan first, and approve or fix it before any code is written.
- Weekly AI Tasks tracker — plan the message→task feature before building; steer if the agent reaches for the wrong storage or skips auth.
- Personal brand site — approve the section outline before it generates prose, so the structure is right before words are spent.
What you learned
Make the agent plan first and fix the approach in words. Then steer as it works — small steps, checkpoints, and “show me first.” It is the build→inspect→decide loop, applied to an agent.