
Why it mattersAn onboarding reader who gets an answer from an agent forgets it the next day. The bet is that a reader who derived the answer under constraints keeps it.
Reading an explanation of a codebase and remembering it the next morning are two different things. Rederive, a new skill for coding agents, takes a different route. It asks only questions, and it will not explain anything.
The project was published five days ago by a developer who goes by EricJujianZou. The GitHub repository has reached 116 stars under an MIT licence. The skill works inside Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, opencode, Windsurf, Amp, Goose and other agents. The author lists git, bash and python3 as the only requirements.
How the session is shaped
Rederive opens by asking why the project exists. On your own repo, it asks you why you built it. On someone else's repo, it works out the original problem itself. It then asks how you would solve it from scratch. While you answer, a background agent reads the source and the import links between files. It also reads the commits where the design actually changed. The author says it does not trust commit messages. A repo built with an AI in a few large commits still works.
The agent builds a private map of five to ten problems. The problems are in the order the system would be built. Each problem carries the constraints that forced the real design. The map lives in .rederive/map.md and holds the answers. A learner is told not to open it until the end. A session log in .rederive/log.md records where the learner struggled.
What happens when you get stuck
When the learner gets stuck, the skill narrows the problem to a few named approaches. The learner picks one. After two wrong answers in a row, it offers to sketch the problem instead. The README shows a shortened session on Express. The user is asked how to handle every HTTP request through one function. They suggest a single handler, then a shared auth function. The skill adds constraints until the user describes a list of functions per route. It then tells them they have rederived a middleware chain.
The author names concrete limits. Everything stays on the user's machine. .rederive/ goes in the repo's local git exclude list, and the skill sends no telemetry. The 116 stars after five days show that developers are starring the repo. The number does not show how many finished a session, or whether they remembered the result afterwards. The method depends on a learner willing to be asked questions for the length of a session.
A new engineer joining a team usually spends the first weeks reading documents. Most of what they read is forgotten by the next morning. Rederive asks the engineer to design the pieces themselves, under the real constraints the project had. The author's claim is that a learner who derived the design will keep it. The experiment is cheap for a team to try. Point the skill at the repository that new hires always struggle with, and compare how the next hire does.
Source
- EricJujianZou, EricJujianZou/rederive on GitHub
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.

