I Didn't Write This

Vibe-coded it? Now understand it.

An agent skill that turns one feature of your JavaScript or TypeScript project into an interactive, evidence-backed lesson: follow the data through the real source, predict what it does, and locate a fault — with every answer that can be checked against an actual recorded execution of your code.

Everything runs locally. No accounts, no cloud, no telemetry.

Install

The quickest way, in any project — the skills.sh installer copies the skill in as plain files you own, for whichever coding agents you use (Claude Code, Codex, Cursor, …):

npx skills@latest add Gjusev/i-didnt-write-this

Non-interactive, e.g. just Claude Code:

npx skills@latest add Gjusev/i-didnt-write-this -s i-didnt-write-this -a claude-code --copy

Requires Node.js >= 22. Two alternatives:

npm install -g i-didnt-write-this   # just the `idwt` CLI (skill ships inside the package)
cp -r skills/i-didnt-write-this ~/.claude/skills/   # manual copy from a clone

Install before starting your agent session, so the skill is discovered on startup.

Claude Code plugin

Claude Code users can install the managed plugin from this repository's marketplace:

/plugin marketplace add Gjusev/i-didnt-write-this
/plugin install i-didnt-write-this@i-didnt-write-this

Use it with your agent

The skill lives in skills/i-didnt-write-this/ (inside the repo, or inside node_modules/i-didnt-write-this/ when installed from npm). How you wire it up depends on your agent:

Claude Code (verified: auto-discovery tested end to end)

Install with the skills.sh installer (see Install) or copy the skill into ~/.claude/skills/ (user-level) or .claude/skills/ (per project) before starting the session, then just ask naturally:

“I didn't write this code — help me understand how the discount logic works in this project, with exercises.”

«No escribí esta funcionalidad; enséñamela con una lección interactiva.»

Claude discovers the skill from its description and runs the whole pipeline: trace, execute in isolated copies, validate, and hand you a lesson.html.

Codex CLI / any agent that reads AGENTS.md

Add a pointer rule to your project's AGENTS.md:

## Understanding existing code
When the user asks to understand, learn, or teach a feature of this
repository (e.g. "help me understand how X works", "I didn't write this"),
follow the instructions in
node_modules/i-didnt-write-this/skills/i-didnt-write-this/SKILL.md
and use its toolkit for evidence, validation, and rendering.

Cursor

Same idea as a project rule — .cursor/rules/i-didnt-write-this.mdc:

---
description: Understand an existing feature via an evidence-backed lesson
globs:
alwaysApply: false
---
When the user asks to understand or learn an existing feature of this
repository, follow
node_modules/i-didnt-write-this/skills/i-didnt-write-this/SKILL.md
and use its toolkit (`idwt`) for snippets, execution evidence, validation,
and rendering.

Gemini CLI

Add the same pointer to your project's GEMINI.md (same content as the AGENTS.md snippet above).

Auto-activation is verified on Claude Code. The pointer pattern works with any agent that reads AGENTS.md / Cursor rules / GEMINI.md — it is the same instruction in three conventions. If your harness has native skills, copy the folder there instead.

Use it without an agent

The toolkit is a plain CLI. Use idwt after npm install -g i-didnt-write-this, or make a project-local install without changing its manifest or lockfile:

npm install --no-save --package-lock=false i-didnt-write-this

npx idwt snippet --repo . --path src/store.js --start 9 --end 17   # verified snippet + hash
npx idwt run     --repo . --entry src/cli.js --input '{"cmd":"x"}'  # isolated run, JSON evidence record
npx idwt run     --repo . --entry __driver.mjs --add __driver.mjs=./driver.mjs   # driver evidence
npx idwt check   lesson.json --repo .                               # contract + freshness validation
npx idwt verify  lesson.json --repo .                               # re-runs all recorded evidence
npx idwt build   lesson.json --repo . --output lesson.html          # standalone HTML lesson

After either installation, invoke the copied skill directly with node <skill>/assets/toolkit.mjs if you prefer not to use the bin wrapper.

Author lesson.json following skills/i-didnt-write-this/references/lesson-contract.md. Run toolkit commands as single node <path> <args> invocations with absolute paths — wrapping them in cd … && … can break permission allowlists.

The resulting lesson.html opens offline in any browser. Answers are self-checked in the page; challenges marked verified show the recorded execution next to them, and lessons whose code has drifted say so instead of guessing.

Development

npm test          # toolkit + fixture tests: contract validation, freshness, executor, escaping, CLI
npm run test:e2e  # Playwright browser suite (challenges, keyboard, escaping, offline, responsive, a11y)
npm run test:all  # both

• fixtures/cart, fixtures/limiter — synthetic teaching projects with known behavior and their own test suites.
• tests/fixtures/*-lesson.json — reference lessons whose recorded evidence is re-executed and compared byte for byte on every test run.
• tests/ — the toolkit's own test suite (plain node:test).
• e2e/ — the browser suite; it builds its lesson pages with the real toolkit pipeline at setup time (@playwright/test + @axe-core/playwright are the only dev dependencies).

What is and isn't verified

• The executor runs .js/.mjs entries with node (custom --runner for TypeScript, e.g. npx tsx; Node >= 23 can also run erasable .ts sources natively). Where code cannot be executed, the skill records challenges as unverified with an explicit note — never as fake evidence.
• The exported HTML renders recorded output only; it never executes code, and all content is HTML-escaped.

License

MIT. See LICENSE.