At the Zo Computer Company, we believe many of the best practices we've learned from coding with AI can be applied to human-AI-computer collaboration in general.
One common practice when coding with AI is to split the work into two phases:
- Planning
- Execution
Planning is where you research the task and formulate a plan. Different people and tools take different approaches for planning with AI. We like to keep it simple. A plan is just a text document, and you can use a Zofile to create a custom mode for writing plans to your liking.
Success in this phase is all about seeding the AI with the right context: relevant files or URLs, any relevant background information you can think of, and your directorial guidance. GPT5, o3, GPT5 mini, and o4 mini are generally good at research and planning, because they're good at reasoning and instruction retention.
It can also be helpful to guide the model to formulate the plan in a certain way. For reference, this is a snippet of our team's planning prompt for coding:
Describe the specific code changes required concisely, with minimal surrounding prose. Design the code changes so that they can be implemented incrementally. Do not break up the changes too much: there should generally be two or three phases of work that logically go together. Any new or changed interfaces in your planned code should be well-typed, self-documenting, and self-consistent with surrounding code.
Execution is where you ask the AI to transform the plan into your desired result. Our preference is to use a fast model for execution. Because we spend a lot of time in the planning phase reviewing and iterating on the plan with our "smart" model, we can use a "fast" model for the execution phase.
In general, there's quite a bit of variation in the "vibe" of each AI model, and you'll find subtle differences in their behavior as you become more familiar with them.