I use ChatGPT projects like small APIs for my repositories

I've been using GPT-6 Astra Pro in ChatGPT on the web to investigate repositories and triage issues without drawing down my Codex usage. In my setup, ChatGPT can read from GitHub and write back when I ask it to. Its task has a hosted working disk for files and tests. The useful part is the project around the chat: I can give it standing instructions and project-only memory, then bring it a specific repository task.

In my testing, I could start 14 requests in a minute. That is an observation from my own use, not a published rate limit or a promise that every task finishes in a minute. It lets me queue several focused investigations while keeping each one tied to a project and repository.

I think of each project as a small API. The instructions define the contract: what to inspect, what evidence to bring back, and when to stop and ask for a decision. A new chat is a request against that contract. GitHub supplies current issues, pull requests, and code; the project carries my standing context.

My setup is short:

1. Create a project for a body of work and select Project-only memory. The setting says the project can access only its own memory, outside chats cannot access it, and Work mode is unavailable for this type of project.
2. Put the standing rules in Instructions. These should describe the work and its evidence standard, not the details of one issue.
3. Select GPT-6 Astra Pro. Connect the GitHub app and give the chat a repository, issue, or pull request link. Check what the agent read and tested before accepting a change or writing back.

For issue triage, I want more than a summary. I want the agent to read the report, find the relevant code, test the likely cause against the source, and tell me whether there is an actionable bug. For development, the same setup can carry work from an issue into a proposed change, with GitHub as the place to inspect and record the result.

One example is my Technical Reconstruction prompt. It asks for primary evidence, labels what is observed or inferred, and works toward a buildable equivalent of a technical system. It is a specialized research instruction, not a general issue-triage template. I keep it in the prompt library so the full text can be read and copied separately from this post.