Here are some tips I have for anyone working with AI or vibe coding right now:
Use dictation
The text to speech microphone on the right of the input box is surprisingly accurate. Force yourself to use dictation when sending new prompts. You can talk a lot faster than you can type, and you typically convey more useful information to the agent about what changes you want to make over voice than text. Don’t worry about your prompt being too long. Remember, agents can see up to 1 million words at a time! A 10 paragraph prompt is a spec in the water compared to the context an agent can see.
For big changes, tell the agent to make a plan and ask you questions
Every time I want to make a major change to a project, I tell the agent to draft me a plan document and put it in a specific folder on my filesystem. This plan helps the agent perform better and makes sure you are on the same page about what is going to be built. It also helps future agents better understand what you built in your repository and acts as a permanent record of documentation about design decisions.
I do not use the native plan tool in Codex or Cursor. Both of these write plans to custom folders on your system that future agents from other providers if you ever switch may have a hard time finding.
I also tell it to “ask me as many questions as you have about implementation details” so that we knock out all of the kinks and logic I did not specify in my original prompt. I keep asking it to ask me questions and respond to my feedback multiple times until I feel the plan is fully fleshed out and the agent and I are on the same page.
Resist temptations to one-shot big ideas. Build in steps by making prototypes first
Starting fresh on a new project or new large feature part of your project? Resist the temptation to give the agent a 3 sentence prompt and hope it interprets what you want to build correctly. Instead, ask the agent to build a mockup or prototype of a user interface or script that does something basic for you to give feedback on. Instruct it not to “wire everything up” or “do a full implementation”. Often times when you ask it to one-shot a full implementation, you may spend more time making changes after the fact than if you had just spent more time planning the implementation first.
Agents have an incredible amount of agency. These are very opinionated creatures that have the ability to build amazing things, just maybe not exactly to your liking. It’s important that you restrain yourself and the agent.
Use the text annotation tool to speed up feedback
Want to respond to a specific part of the response an agent just gave? Select the text in the chat window and click “Add to chat” in Cursor or Codex. You can then comment on that specific part of the prompt. You can also do this for documents open in the side panel of Codex or Cursor.
Make accurate user interface changes with Design Mode
Cursor and Codex both have an integrated Browser that allows you to open the website you’re vibe coding on the right and point out / or drag highlight areas that you want the agent to change. This allows you to direct the agent to change specific parts without you having to bother typing out what part you want changed. Here are the docs:
Cursor: https://cursor.com/blog/design-mode
Codex: https://learn.chatgpt.com/docs/browser?surface=app#app-comment-on-the-page
I highly recommend trying it out!
Use faster models
If you have Cursor, Composer 2.5 (even with fast mode turned off) is incredibly fast and useful. GPT 5.6 Luna in Cursor and Codex is a smaller and faster model too.