Getting Started with Plan Mode
A practical guide to using AI planning modes effectively. Stop treating your agent like a magic wand and start treating it like a junior engineer.
2025-11-05
Loading content...
A practical guide to using AI planning modes effectively. Stop treating your agent like a magic wand and start treating it like a junior engineer.
Skipping the planning phase with AI agents leads to hallucinated architectures and overwritten code because the model codes before it thinks. By forcing the AI into "Plan Mode" to outline its strategy, draw diagrams, and define tests first, you catch misunderstandings before they become technical debt.
We have all been there. You describe a complex feature to an AI, hit enter, and watch in horror as it immediately starts over-writing your critical AuthService with a hallucinated version of passport.js from 2019.
The solution isn't to write better prompts. The solution is to stop letting the AI code before it thinks.
Most modern AI coding tools now have a "Plan Mode" (or "Composer" or "Architect" mode). This is a dedicated state where the AI parses your request, reads your files, and proposes a strategy before it touches a single line of code.
If you aren't using it, you are coding with one hand tied behind your back. Here are the core rules for getting started.
Here is my rule of thumb: If you wouldn't trust a new intern to pick up a Jira ticket and complete the task in one go without supervision, you probably shouldn't trust an AI to do it either.
Complex tasks require supervision. Plan mode is that supervision. It forces the AI to show its work, allowing you to catch architectural misunderstandings before they become technical debt.
We are trained to be concise. "Fix bug." "Update CSS." "Refactor utils."
In Plan Mode, this is the wrong instinct. You have a context window of 200,000+ tokens. Do you know how much text that is. That is roughly the length of A Court of Thorns and Roses. You are not going to overflow the buffer by chatting.
Be messy. Be verbose. Paste in your rough notes, your half-baked thoughts, and your worries. The more context you provide about why you are making a change, the better the plan will be.
Most engineers can talk significantly faster than they can type. Modern AI tools are surprisingly good at transcribing technical jargon.
Hit the microphone button and just ramble for two minutes. "I want to add a user profile page, but I'm worried about how it interacts with the existing session management. We need to make sure the avatar upload doesn't block the main thread..."
This stream-of-consciousness input often contains subtle constraints that you would edit out of a written prompt.
Your AI doesn't know about that internal wiki page you wrote last week. It doesn't know about the updated API docs for that third-party library you just installed.
If you have relevant documentation, copy it directly into the plan. Don't assume the AI knows it. Treat Plan Mode as a shared workspace where you dump all the puzzle pieces before asking the AI to assemble them.
Text is often ambiguous. A "flow chart" is not.
One of my favorite tricks is to ask the AI to generate a Mermaid diagram of its proposed plan.
"Draw a sequence diagram used mermaid.js showing how the User, the API, and the Database will interact during this transaction."
Seeing the logic mapped out visually often reveals race conditions or circular dependencies that are invisible in a bulleted list.
The first plan is rarely the final plan.
Review the AI's proposal. Challenge it. "You missed the error handling for the network timeout." "This feels over-engineered, can we simplify the state management?"
This iteration cycle is where you save time. Spending 5 minutes arguing with the AI in Plan Mode saves you 5 hours of debugging a broken implementation later.
A plan shouldn't just include how to build the thing; it should include how to prove it works.
Talk to the AI about the "completed state."
Make the validation steps explicit requirements of the plan.
Finally, a critical technical tip: Clear your context before you start planning.
If you have been debugging a different issue for the last hour, your AI's context window is full of irrelevant noise. You don't want that Context Poisoning leaking into your new feature.
Or, even better:
This gives you a clean slate with zero baggage, ensuring your agent is focused entirely on the execution of the agreed-upon plan. Saving your context in a document is the ultimate way to maintain focus.