Analysis updated 2026-05-18
Compare how safely an AI assistant implements a ticket with and without extra project context.
Use a hidden PHP safety checklist to grade an AI generated implementation.
Study reference snapshots of safe versus unsafe implementations of the same feature.
| ok-up/domain-pack-workshop | boundwize/structarmed | mugisha-eric-250/schoollibrarymanagementsystem | |
|---|---|---|---|
| Stars | 13 | 14 | 12 |
| Language | PHP | PHP | PHP |
| Setup difficulty | easy | easy | moderate |
| Complexity | 2/5 | 3/5 | 3/5 |
| Audience | developer | developer | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires PHP 8.3 or newer and Claude Code available on the command line.
This project is a workshop style experiment for testing whether an AI coding assistant implements a feature ticket safely, and whether giving it extra project context helps it avoid mistakes. It is built around a small demo project called meetly with a blocking feature that needs to be added. The repository sets up two copies of the same starting project. One copy only has a plain task description file telling the assistant what to build. The other copy adds a CLAUDE.md file and a set of domain pack files that describe rules and context about how the feature should behave. Both copies are turned into separate git repositories with no remote, so changes can be tested and reset independently. The workshop also includes two reference snapshots showing what happened after running the experiment: one snapshot shows the result from the plain project, described as a less safe implementation, and the other shows the result from the project with the added context, described as a safer implementation. A hidden safety checklist, written as a PHP test file, is used to grade whichever implementation the assistant produces after the fact. To run it, you need PHP 8.3 or newer, Claude Code available on the command line, and git. A setup script creates the two independent project copies. You first confirm a baseline test suite passes, then ask Claude to read the task file and implement the ticket. Afterward you copy in the hidden safety checklist and rerun the tests to see how many pass, comparing the result against the provided reference snapshots. The same steps are then repeated for the version with the context pack, to compare outcomes. A final cleanup step resets both project copies back to their original state.
A workshop that tests whether giving an AI coding assistant extra project context helps it implement a feature ticket more safely.
Mainly PHP. The stack also includes PHP, Claude Code, git.
Setup difficulty is rated easy, with roughly 30min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Verify against the repo before relying on details.