← Back to Portfolio
Personal · project

Shikhu — learn the code your agents write

A CLI and agent skill that quizzes you on AI-generated code and tracks how much of your codebase you actually understand.

Shikhu is a command-line tool and agent skill I built to close the gap between generating code with an agent and actually understanding it. It reads your codebase file by file, generates conceptual quizzes, and tracks how well you do over time — a metric I call knowledge coverage, kind of like test coverage, but for your own understanding.

How it works

  • Generates multiple-choice questions from your source files using Mercury, a fast text-diffusion model
  • Tracks "golden" questions — ones you answer correctly and mark as good — with three per file counting as full coverage
  • Detects when a file changes and re-validates the questions that have gone stale
  • Ships a /shikhu-study skill that walks you through a file and seeds new quizzes from the questions you actually ask

Why I built it

Recent research suggests coding agents can quietly erode the very skills we need to supervise them. Shikhu is my attempt to keep learning while still moving fast — by making conceptual engagement a measurable, repeatable habit. Everything lives in a local SQLite file; nothing leaves your machine except calls to the Mercury API under your own key.

Skills & Technologies

Python CLI LLMs Agent Skills SQLite Developer Tools Mercury / Inception
// stay in touch

Get Answering Machines in your inbox.

Occasional notes on AI, coding agents, and where this is all headed — plus the very occasional origami diagram.

Arjun Kirti Patel
Chicago, IL · folded with care
© 2026 Arjun Kirti Patel · MIT Licensed