How to (really) use Plan Mode
Spend some time before, or waste more time after, your pick.
Because providers insist so much on using “thinking” terms across the agents interface, people tend to forget how LLMs work.
All LLMs really do is predicting the next word they will generate, from the current context, plus its general knowledge.
That context includes:
The current conversation (ie. your brief)
Other knowledge it has about you
The existing codebase
Plan mode forces the agent to ask you questions about gaps in your brief and generate implementation instructions (how it should be done)
So when it gets to generating the actual code, the implementation instructions it wrote for itself become part of the context. That has a huge impact on the quality of the code generated.
Now, here’s how you go further:
When writing your prompt, make sure you ask the agent to:
Confirm understanding and assumptions from your brief
Critique your approach/suggest alternatives
I know it feels like a waste of time (and tokens) because “it’s not that complex of a task” and “isn’t plan mode already doing that anyway?”
Well, yes and no. Something else to remember:
An agent’s primary job is to make you happy, instantly (so you keep using it)
If you think about it, questioning everything somebody says is not necessarily the best way to make them happy.
You’d better off doing what they say, and do it now, using whatever assumption you extracted from their brief.
Now, if they asked you to be critical, you can take a step back, think about the bigger picture and do the “right” thing. It’s the same with your agent. Maybe even worse.
By using Plan mode like this, you will be surprised how much context, assumptions and decisions only lived your head until you start getting the questions :)