Guide · Updated
Tools that check you understand AI-written code, compared
Quiz extensions, pull request gates, Claude Code learning modes, AI review bots and attribution tools: what each checks, which ones block, and where TypeBack fits.
Plenty of tools now promise to make sure someone understands the code an agent wrote. They answer different questions, and several of them work by blocking something: a commit, a push or a merge. This page sorts them into four kinds, says what each is good at, and where TypeBack fits. We make TypeBack, so weigh our view of it accordingly. Descriptions are from each tool's own listing, checked in September 2026.
Four kinds of tool
| Kind | Question it answers | Examples |
|---|---|---|
| Comprehension checks | Does the person understand this change? | AIditor, Reviewsaur, Gater, TypeBack |
| Learning modes | Can the agent teach while it works? | Claude Code's Explanatory and Learning output styles |
| AI review bots | Is this code any good? | CodeRabbit, Qodo, Greptile, Bito |
| Attribution | Who, or what, wrote this line? | git-ai, Agent Trace, Co-authored-by and Assisted-by trailers |
Only the first kind checks the person. The other three are useful, and a team may want all of them, but none tells you whether the author could explain the change.
Comprehension checks
These ask the developer questions about AI-written code. They differ in where they run and whether they block.
| Tool | Form | What it does |
|---|---|---|
| AIditor | VS Code extension | Quizzes you on staged changes; installs a pre-commit hook that blocks the commit until you pass. |
| Vibe Check | Open VSX extension | Turns AI-written code into a Duolingo-style quiz, with spaced repetition. |
| Atomic Reps | Open VSX extension | One short question about code you just wrote, at a natural pause. |
| CopyWorkCode | Open VSX extension | Review AI-generated changes by typing them yourself. |
| Waitso | Open VSX extension | Quizzes and short lessons while your agent works. |
| Contral | VS Code extension, paid | Explains every edit an agent makes, as a teaching assistant. |
| quiz-me, learning-with-claude | Claude Code skill and plugin | Quiz you on your own diffs inside Claude Code; quiz-me can add a pre-push hook. |
| Commit Comprehension Gate, PR Quiz | GitHub Actions | Ask the pull request author questions; the merge waits until they answer correctly. |
| Reviewsaur, Gater | Hosted GitHub apps | A quiz before a pull request can be approved or merged. Reviewsaur lists $11 per seat a month, billed yearly. |
Three questions separate them:
- When does it ask? At the edit, at the commit, at the push, or at the pull request. The later it asks, the less you remember of what you asked the agent for.
- Does it block? Gates are simple to reason about, and people route around them. When a pull request quiz launched on Show HN in 2025, a common objection was that developers would have the AI answer the AI's quiz.
- Where does your code go? Hosted apps read your repository. Editor extensions can use the model you already have, including a local one.
Learning modes
Claude Code has two output styles aimed at understanding. Explanatory adds short notes on why Claude made each choice. Learning asks you to write small pieces yourself and leaves TODO(human) markers. Anthropic's skill-formation study points to modes like these as a way to keep learning while using AI.
They change how the agent talks during the session. They leave no record of what you understood, only exist in one agent, and give a team nothing to check. They pair well with a comprehension check rather than replacing one.
AI review bots
CodeRabbit, Qodo, Greptile and Bito review the code: bugs, style, security and missing tests. They are good at catching what a tired reviewer misses. They say nothing about the author. A pull request can pass every bot and still be code nobody on the team can explain, which is the comprehension debt problem.
Attribution
git-ai records, line by line, which code an agent wrote, in git notes. Cursor's Agent Trace proposes a format for linking lines to conversations. Trailers like Co-authored-by and the Linux kernel's Assisted-by mark a whole commit. All of them answer “who wrote it”. None answers “did anyone understand it”.
Where TypeBack fits
TypeBack is a comprehension check that runs in VS Code. Its choices, and what they cost:
- It asks at the edit. Hooks for Claude Code, Copilot agent mode, Codex and Pi queue each hunk as the agent writes it, so you can review while the agent keeps working. Cost: it needs the extension and the hooks installed.
- It never blocks. You can skip any hunk, and the skip is recorded. Cost: if you want a hard gate, the only one is the optional pull request check.
- It asks more than recall. Predict an output, trace a path, find a planted bug, or say why the code is written that way. There is a typing drill too, which CopyWorkCode also offers.
- It keeps a record in the commit. A
Reviewed-by-humantrailer, and for teams a record the freetypeback/verifycheck reads. It is the author's attestation, not proof. - It has no server. Questions come from the model you choose, including a local one. No telemetry. Cost: question quality depends on that model.
How to choose
- You want a hard stop before every commit, and you work alone: a pre-commit quiz such as AIditor.
- You want to learn while the agent explains itself, in Claude Code: turn on the Learning or Explanatory output style.
- You want a second pair of eyes on code quality: an AI review bot.
- You need to know which lines an agent wrote, for audit or licensing: git-ai or a trailer convention.
- You want to understand each change as it is written, never be blocked in the editor, and give your team a record it can check: TypeBack.