been using ChatGPT and Codex combo for a while. Figured i’d share my method i’ve fallen into.I am not a programer and mostly just seeing what I could do.I started with Chat, talking ideas about what i wanted the project to be. What i wanted the user experience and workflow to behave like and act. Worked though mostly the high level aspects of it all and distilled it a narrative of the purpose of the project.I used Chat as my design and research thread, to figure out what was out there in the ecosystem for python, what to use/pro/con eval.
From there, i had Chat start writing prompts for codex to implement.I didn’t tell it how to write the prompt or how big the slice should be, other than go slowly. Only told it 2 things: use 1 non-nested triple backtick block for easy copy pasta and to keep the unrelated / negative instructions to a minimum.
Chat would conjure up 200 line prompts for codex to work from. Reading though each one myself (learning and watching for drift), the prompt would describe the boundaries and aspects of the slice, tests it wanted made and run, and the report that it wanted back from codex about how it did things.
I’d copy the prompt into Codex (VS Code Extension) and run it. Wait for it to finish and copy the midstream thinking and it’s final report back into ChatGPT. I’d work with chat to reason about the next slice, what it should be and if there were any parts that needed product decisions that needed to be made (think /grill-me skill but by hand), and move on to the next prompt. Periodically, and depending on what slice Chat was on, it would write out an audit prompt to inspect the repo for implementation details to plan on/around/fix.
This went on for months, I was just mostly seeing what i could do with it.
I came a way with a few sharp edges i fell on…
I found that around 600 lines of code in a file the apply_patch tool starts to choke or something. When it did, codex would inhale the file into context, delete it, then exhale the same file out just to change 5 lines of code.So i came up with a 300 line Smell Rule. At 300 lines in a file, i would have Chat/Codex evaluate and see if the file had too much responsibility and split along reasonable and useful breakpoints.
Naturally this lead to file sprawl. I had Chat/Codex start to organize files by ownership layers, as many folders deep as was was useful and meaningful. I had it keep files clustered by responsibility, (the basic of my project was very CRUD, so things fell naturally together) . File names had to be descriptive of responsibility, with a Role-based-suffix. In the agent.md I spelled out exactly what suffixes to use and what they meant, and it added a few as the repo grew.
This had the side effect of turning a file’s path and name into compressed semantic hints when it went searching through the repo. No, i have no MCP of any kind loaded and only in the last month did load up any skills: Thermo-Nuclear Code Quality Review by Cursor, improve-codebase-architecture skill from the mattpocock, deslop & signage. Between the 2 code review skills, most of what it kept finding were slightly fuzzy contracts and stylistic differences of opinion.
Repo Documentation – have the agent document what it’s creating. APIs / Architecture / UI Style Guides. If it’s on a stretch, let it keep a diary. Once it’s done with the stretch, wipe the diary and re-document from code itself and evaluate if that’s what was intended in the first place. deslop skill comes in handy here , other times you want the obviously AI level of explicit detail. Sometimes it’s also necessary to make a set of constitutional documents that exactly and precisely spell out in every detail how and why something is the way it is and should remain so. Prevents the agent from re-litigating from generic patterns that don’t apply and are not valid for the implantation. For me, I was making an app for the self-hoster/home lab target group, didn’t need SaaS/Ent level mechanics and it kept trying to cargo cult it in half way.Stay away from the Jargon. It is too much subtext and baggage that LLM’s simply don’t get. Ask 5 people in any profession a debatable question, you’ll come away with completely different answers that are all technically correct and completely specific to the situation but still all the same thing. “What does DDD actually mean down in the repo?”
I didn’t really know that context window compaction caused issues… i just let codex do that on it’s own, never really noticed anything good,bad, indifferent about it. Chat certainly doesn’t let you control that or even tell you it happened. When i did come across it in YouTube videos, the whole “smart zone” vs “dumb zone” thing really threw red flags for me all over the place. Made me wonder what was going on that caused either. Agent systems to me seem to be like really smart dumb people. Very good at following directions if you give them a decision domain that is properly bounded and explained. Otherwise they seem to suffer the same problem anyone has with too much information in front of them trying to make a choice. Both cases they start to hallucinate.
--Thought i’d throw my 2c into the ringo/
