Hot Tips for AI-Powered Coding

Here is some hard-won knowledge I’ve picked up to get more consistent results during AI-assisted development. Most of these are just applying basic software engineering principles to vibe coding, but I don’t hear them talked about enough.

Weathered textile under tension with cinematic lighting and subtle purple-pink hues, representing dynamic balance.

Initial Setup

Make a docker container for your development environment. Make a README.md, docs/ROADMAP.md, docs/frameworks/{your_framework}.md, and then an AGENTS.md that references these.

Mandatory Planning Stages

For the initial buildout then for every major feature, make a well-defined REQUEST to answer all the relevant questions beforehand. There’s most likely a lot that you’re not thinking of.

Turn the request into a SPEC and turn the spec into a PLAN, then put both the spec and plan into docs/features/{feature_name}.md.

Version Control

Use git liberally, it’s great. Do your work on feature branches, use the staging area as you work, and make lots of commits in case the model does something undesirable.

Also make use of GitHub. Start pushing your feature branches and making PRs for every major feature or step. Install the GitHub Pull Requests extension or equivalent so you don’t have to leave your editor.

Automated Testing

One thing LLMs are exceptional at is generating tests. Do some TDD (here are instructions for python and typescript) or at least incorporate testing into your workflow. Not only will you be so much safer making changes, but any good agent is going to be running that test suite and if you can get them in a good testing and coding feedback loop then you can chill.

Code Review

You will catch a lot if you have Copilot do a review on your PR in GitHub or use something like Code Rabbit. You can catch some more with a REVIEW prompt. Still there’s really no substitute for reading the code yourself and going over the test suite to know its intended behavior.

Parallelization

Figure out parallelization that works for your setup. Browser agents like Codex and Jules make this easy. Some people use git worktrees and multiple IDE windows. The future is managing a team of agents, breaking plans up by which parts can be run in parallel, and mostly just reviewing a lot of PRs. Plus this way you can give the model multiple tries to complete the task and pick the best one.

Conclusion

By following these practices, the variance between models will be reduced and changes will be safer to make. The plan provides the guardrails to keep them on track, and the rigor around testing and version control provide the safety net.