Do-I-understand – Check you understand AI-written code before you merge
Reverse code review flips the paradigm — AI quizzes YOU before merge.
Learn your codebases through self-quizzing and study. Track your knowledge coverage over time.
Knowledge coverage metric tracks how many files you actually understand, not just generated.
Developers using AI coding agents who want to retain conceptual understanding
Cursor · Continue · Sourcegraph Cody
I'm Arjun, I'm currently a developer advocate at Pinecone, and I use agents to code a lot!
I enjoy using agents to code, but I've been feeling like I've lost some conceptual learning and understanding that comes with writing the code yourself. I wanted to make a tool that would make it easy to re-build habits around learning, while complementing workflows that use agents. I found inspiration for the tool from reading a research paper on skill formation from Anthropic, and Shikhu is the result of my attempt to build a tool around that: https://arxiv.org/abs/2601.20245
Shikhu focuses on helping develop conceptual understanding via knowledge coverage on a file-level.
Shikhu makes multiple choice quizzes on a file-level for you to take. You can reject a malformed question, answer it, and eventually create validated "golden" question sets for your code. Your ability (or, anyone who you collaborate with) to answer golden questions about your code corresponds to your knowledge coverage. Using the skill, you can learn specific files with your own coding agents, and use those conversations to inform quiz development, to help reinforce learnings.
Repo is here, it's free to use, and requires an Inception API Key for question generation and summarization:https://github.com/arjunpatel7/shikhu
Install Shikhu with uv: uv tool install shikhu
Install the agent skill afterwards:
shikhu init shikhu install-skill
As of writing, the free tier of Inception's pricing should cover typical use.
My article where I explain more about Shikhu is here: https://www.arjunkirtipatel.com/blog/introducing-shikhu
It's definitely rough around the edges, probably will change a lot and maybe has a few bugs, but I found it useful enough to share. Thanks for reading!
-Arjun
Reverse code review flips the paradigm — AI quizzes YOU before merge.
Local codebase model for AI agents when Cursor and Sourcegraph already index code.
Persistent context layer beats Cursor's session amnesia on large codebases.
Auto-generates multi-tool agent configs from codebase introspection—fills real AI coding friction.
Visual knowledge graph navigation beats pure Q&A, but Cursor already exists.
AST + embeddings for codebase search—but Sourcegraph Cody, Cursor, and Continue already solve this.