Updating a hackathon README for an attribution rule: derive credits from the manifests, separate "ours" from "not ours", and do not write claims you cannot verify
Task: a team needed their README updated to satisfy an event rule: credit every public framework and library, clearly distinguish your own contributions, be able to explain your code, and do not overstate contributions (including LLM use).
What worked:
1. Build the credits from the dependency manifests, not from the existing README. Reading package.json and pyproject.toml found one library the README credited that was not a dependency at all (a numeric library; the code in question was plain Python) and four that were used but uncredited (a second LLM adapter, a vendor SDK, a QR-code package, and the audit tools mentioned elsewhere in the README). An over-credit is as wrong as an under-credit when the rule is about accuracy.
2. Replace a flat "Built with" list with three parts: - Our work: named by directory, one line each on what was actually authored (rule engine, adapters, geometry code, UI). - Not our work: frameworks and libraries (linked), data sources and third-party services, embedded third-party UI (a vendor viewer and its tools), and the published standard whose thresholds were implemented. People forget the last three categories. - AI assistance: a plain statement of how assistants were used.
3. Ground statements in things you can check: git log --reverse for the first commit date, git log --format=%an | sort | uniq -c for authors, and a grep of commit trailers for assisted commits. I first wrote "everything was written during the event" and cut it back to "the history starts on <date>", because I could not verify the stronger claim for every file.
4. For the AI paragraph, write only what you know and hand it back. I could vouch for the commits I co-wrote (they carry a co-author trailer) but not for how other teammates worked, so the paragraph says assistants were used under the team's direction and I told the user plainly that every teammate has to be able to stand behind it and should edit it. Do not write "the team wrote all the logic by hand" or similar on their behalf.
5. Update figures that have gone stale (test counts, sample sizes replaced by a full survey) in the same pass, since a README that contradicts the code undermines the attribution section next to it.
6. Check the rule for other surfaces. This one required the same credits on the submission page as well as the README.
Also: one line of the forwarded request was ambiguous ("get rid of" two section references that appear in the README, the UI text and the rule engine). I listed the readings and asked, instead of deleting something central on a guess.