What to decide before handing work to AI
I grouped the work rules actually written in the Strayling, Nook and this site's repositories into five. No made-up examples, and I noted where each rule is written.
What not to do
At the start, I write that anything hard to undo or that costs money, like deploys, domain setup, making things public or payments, does not happen without approval.
- Strayling's CLAUDE.md: Claude starts a deploy only when asked.
- Strayling's CLAUDE.md: server secrets are not committed to the repository.
- This site's work instructions: do not push to main directly, and do not use destructive git commands.
Report format
I get results in a form I can check, not just words.
- PR titles and descriptions are written in Korean, and say what changed and how it was checked.
- Run the checks yourself and put the commands and pass counts in the PR description.
- This site keeps screenshots at 1440 and 390.
Roles
Split who builds, who reviews and who decides. For this site, one session builds and another reviews, and merging happens after review. Things a person has to do are gathered as a checklist in one issue, as in Strayling.
Do not make things up
I write that posts that do not exist stay as coming soon, and numbers and examples are not filled in. For Strayling's store copy, each line has a table entry pointing to the code file it is based on. This site's app pages also have a commit that removed made-up request sentences.
Source documents
Strayling sets its planning baseline in one file, PLAN.md, and keeps the rules the code enforces in PRODUCT.md. If they differ, PLAN wins. But docs can fall behind the code. When I read the repositories this time, there were a few places where the docs did not match the code.
Check rules
- Nook: keep lint warnings at 0 and do not add new ignore comments.
- Nook: run frontend lint, type checks and backend tests on every PR.
- Strayling: run the same checks on the Mac without CI, and the deploy script runs them again before building the image.
When stuck
If a decision is needed, go ahead with the recommended option first and report it separately. The work does not stop, and the decision stays with me. Anything hard to undo or that costs money is the exception, and work stops there.