This is an experiment to see if I was able to reproduce Harper Reed's process.

I created the spec and the plan with ChatGPT, then developed initially with Claude Code. When I started hitting the 5-hours limit (I'm on the $20 plan), I used other tools. I used briefly Gemini CLI, then ONA, then Codex extensively.

Notable points

• The acceptance tests are the strongest guardrails. They provide assurance that the code does what it's supposed to do in a clear way (check out the acceptance tests). Unit tests are useful too, as design feedback to the AI, but they are not as straightforward to use for checking regressions and progress
• There is exactly one commit where I wrote code with my own hands, and it was where I was trying to get the AI to simplify the code. The actual idea was suggested by Claude in a conversation, where he said that error line tracking is only needed during parsing, not during program execution. I tried to get both Claude Sonnet 4 and GPT 5 to make the simple change to track the file line number in the ParseProgram method, but wasn't able to make myself understood. In 10 minutes of human code, I made the change that proved that the line number tracking in the AST nodes wasn't needed, and then got Gemini to clean up the dead code.
• Alternate step implementation chats with refactoring chats. Discuss refactoring options with the AI. Avail the awesome Kent Beck Code Mentor agent by Giordano Scalzo.
• Keep the project documentation up to date. Follow the advice from Shrivu.

Status: mostly done. I don't expect to do much work on this anymore

Check out classic basic programs such as

scripts/run.sh testdata/guess_number.bas

or

scripts/run.sh testdata/hamurabi.bas
scripts/run.sh -max-steps 100000 testdata/wumpus.bas