Full honesty up front, because this repo demands it.

rederive was created on October 19, 2026. When I found it, it had 24 stars; it was already at 71 stars a day later, as of October 2026 — which still makes it a tiny, days-old, single-author project that is essentially untested by the crowd. I’m not hyping a proven tool. I’m showing you an idea early, because the idea is the point.

Every other tool in Repo Week made the AI stronger. This one makes you stronger. rederive is a skill that teaches you a codebase by making you re-derive its design decisions yourself — and it will not, under any circumstances, hand you the answer.

If you’ve ever vibe-coded a build you can’t explain, this finale is aimed directly at you.

The problem: reading code isn’t how you learn a codebase with AI

Here’s the standard move now: point an AI at a codebase, ask it to explain the architecture, read the summary, nod along. You feel informed. You are not — you’ve memorized a description of understanding, not the understanding itself. Ask follow-up questions a week later and watch how much of it evaporates.

rederive is built on the opposite bet, and it names its target: Bloom’s creation layer. In learning science, creation sits above recall and explanation — you learn by building the reasoning yourself, not by consuming somebody else’s. Reading code is remembering. Rebuilding its design decisions, decision by decision, is learning.

So rederive flips the tutoring model. Instead of the agent explaining and you nodding, the agent withholds — deliberately, by design — and you do the deriving. The agent holds back so your brain does the work. That’s augment, don’t replace, applied to education: the tool’s job isn’t to think for you, it’s to make your thinking unavoidable.

How rederive works: map first, then problems in build order

Here’s a session, step by step, from the source material.

It starts from why. The process begins from why the project exists — before any file gets opened, the point of the codebase frames everything after it.

A background agent maps the repo. While you start working, an agent maps the codebase for you: the files, the import links between them, and the commits where design decisions happened. That map becomes 5 to 10 problems, in build order — the sequence in which the project was actually constructed, reconstructed as exercises.

You solve each one. Real constraints get added one at a time, the way they arrived in the real project. You derive the design instead of being told it.

Stuck? It narrows. It never tells. When you’re stuck, rederive shrinks the problem down to a few approaches. Get it wrong twice, and it offers to sketch the problem — still not the answer. The sketch points; you still walk.

Then it shows you the truth. The signature moment is the Express example from the source material: run it on Express and you end up inventing the middleware chain yourself — and then it shows you that Express did the same. You didn’t read about middleware. You needed it, invented it, and then met the original. That lesson doesn’t evaporate in a week.

Everything stays local, too: the map and your notes live in a .rederive folder — the map in .rederive/map, your notes in .rederive/log.md — git-excluded, never committed, no telemetry.

How to install it and run your first session

Installation is one line:

npx skills add EricJujianZou/rederive

Add -g to use it in all projects. Then run it inside your agent — the example invocation from the source material:

/rederive expressjs/express

It runs in Claude Code, and per the source material it also works in Codex, Cursor, and 20+ other agents. Requirements are modest: git, bash, and python3.

Required caveat: read the skill files first

This section is not optional, and rederive’s own framing insists on it.

The repo is days old, single-author, and untested by the crowd. Small and new isn’t a sin — every proven tool started there — but it changes your obligations. Read the skill files before you install. Understand what it will do on your machine before it touches your machine. That’s true for any repo you pull off the internet, and it’s doubly true for one that’s a week old.

Judge rederive the way its own post frames it: as an idea worth steelmanning, not a tool with a track record. If the idea is right, the star count will take care of itself. Judge it again in a year.

Who should learn a codebase with AI this way

Anyone who builds with agents and has felt the specific shame of shipping code they can’t walk through. The vibe-coded app that works — until it breaks, and you realize you can’t reason about your own system because you never built the reasoning in the first place.

Also: engineers picking up an unfamiliar legacy codebase, students who keep “learning” frameworks by watching instead of deriving, and anyone who suspects their AI workflow is quietly deskilling them. rederive is the counterweight — same agents, same era, opposite direction. The whole Repo Week series, including the tools that make the AI stronger, is rounded up here: https://bigjermsays.com/how-to-make-money-with-ai-build-a-system-not-a-side-hustle/; the systems I build on these ideas live at https://bigjermsays.com/shop/.

The repo, in detail

Repository: github.com/EricJujianZou/rederive
License: MIT
Stars: 24 when first checked, 71 a day later, as of October 2026 — tiny either way
Created: October 19, 2026 · single author

What it is: a skill for Claude Code and 20+ other agents that teaches you a codebase by withholding the explanation. A background agent maps the repo — files, import links, and the commits where decisions happened — and converts the design into 5 to 10 problems in build order. You solve them; hints narrow the problem when you’re stuck, a sketch is offered after two wrong answers, and the answer itself is never given. Map and notes stay local in a git-excluded .rederive folder, with no telemetry.

Who it’s for: builders who want to understand what they ship — especially anyone re-deriving a build they vibe-coded but can’t explain. Who it’s not for: anyone who needs the explanation fast for a production incident. This is a learning tool with a deliberate pace, not a documentation generator.

How you’d use it: install with the command above, run /rederive against a repo you want to understand — yours or a famous one like Express — and work the problems in order. Read the skill files before installing, because it is brand new.

FAQ

Is rederive a proven tool?

No — and it doesn’t claim to be. It was created on October 19, 2026, is single-author, and sat at dozens of stars as of October 2026. It’s an early idea presented honestly, not a battle-tested standard. Read the skill files before installing.

Why won’t it just explain the code? Isn’t that slower?

Deliberately, yes. Explanation produces recognition; derivation produces understanding. rederive targets Bloom’s creation layer — you reconstruct the design decisions yourself, with narrowing hints when stuck — so the knowledge survives contact with a real problem later.

Does rederive send my code anywhere?

Per the project, no telemetry: the repo map and your session notes are stored locally in a .rederive folder that is git-excluded and never committed. As always with a brand-new tool, verify that in the skill files yourself before running it on anything sensitive.

The bottom line

Six repos this week made your AI faster, smarter, and better dressed. The finale is the one that refuses to do the work for you — and that’s why it’s the best idea of the week. Go re-derive a build you vibe-coded but can’t explain. Save this post — and the other six — because together, they were the whole starter kit.

Big Jerm