Why Vouch is like this

What was broken

You asked an AI for a feature. It worked. You merged it and moved on. Three months later it breaks at 2am, and you are reading code with your name on the commit as though a stranger wrote it. You do not know why that package is there. You do not know what happens when that call fails.

That gap stays invisible until it is expensive: in an outage, in a review, in an interview, or the moment somebody asks why did you do it this way and you discover you do not know. Nothing in the toolchain surfaces it, because compilers and linters check the code, not the author.

And the tools that claim to help make it worse. They summarise your codebase for you, and reading a good summary produces the confident feeling of knowledge with none of the ability to produce it.

Who it’s for

• People shipping fast with AI who suspect they are getting worse, and want to know for certain
• People who inherited a codebase, including the ones who wrote it and cannot tell the difference any more
• People going into interviews who want their real gaps, not their imagined ones
• People on call for code they skim-read

Who it isn’t for

• Anyone who wants a tool that reviews or explains their code. That is the opposite of this
• Teams looking for a quality gate in CI. It never blocks anything
• Anyone unwilling to be told they were wrong about themselves

Rules we don’t break

Decided once, so they never have to be argued about again.

• Ask before you explain. You rate your confidence and attempt the answer first. The real answer is withheld until you do, and that is enforced in code and covered by a test, because it is the entire mechanism.
• Never block, never gate. Not a linter, not a reviewer, not a CI check. The Claude Code hook cannot fail your session: an exit code that would erase what you just typed is never used, and every error path exits silently.
• Never get between you and AI. It assumes you will keep using AI and is designed for that world rather than against it.
• One number that is hard to argue with. The Gap: confidence minus what you demonstrated, per area of the repo. Specific, personal, and the only score here that matters.
• Your code stays yours. Redaction runs before every model call with a never-send list that cannot be overridden, hunks redacted past 40% are dropped entirely, and the outbound host list is fixed in code. vouch purge deletes everything.
• Earn the install in two minutes. The first command needs no AI, no key and no account, and pays for itself by deleting packages you were shipping for nothing.

What it actually does

• vouch unused. Packages nothing in your source imports, with the ones that legitimately have no import site kept in their own list so they never inflate the number. On a real Next.js app: 16 findings, 10.7 MB, 16 for 16 confirmed by hand.
• vouch digest. After a work session: five items maximum, one real question each. Confidence first, answer withheld until you attempt it, then the reveal with install size, licence, known vulnerabilities and what a paid service costs at scale.
• vouch defend. Shows you only the filenames from a change you shipped and asks you to reconstruct what it does and what it assumes. Then it reveals the brief written when the code landed.
• vouch map, cards and trend. A treemap of your repo coloured by what you can defend, with a shareable PNG. Free instant re-tests of things you learned weeks ago, which visibly decay. And your progress over weeks.
• vouch init and status. Choose per area whether you are staying sharp or deliberately outsourcing it, then see where you stand and the one useful next thing.
• The Claude Code plugin. Before Claude answers a substantive question in a repo you are keeping sharp, it asks you to predict the shape of the answer. Sampled and cooled down so it stays rare, always skippable, switchable off entirely.

What it refuses to do

The most useful part of any note like this. Every line has a reason attached.

• Reviewing or explaining your code. The category is full of these, and they are the reason the gap exists. Producing the answer is the product.
• Being a CI gate. Anything that can fail a build becomes something to route around, and this only works if you are honest with it.
• A dashboard for your manager. The Gap is a personal number. Reported upwards, everybody would lie to it immediately.
• Sending your code anywhere we control. No account and no server of ours. Redacted excerpts go to Anthropic through your own CLI; package names go to public registries for licence and vulnerability data.

How we’ll know it worked

• The Gap shrinks for a returning user, in the areas they chose to stay sharp in.
• People get questions wrong and keep running digests anyway. A tool that only ever confirms you were right is measuring nothing.
• vouch unused finds something real on the first run, which is what buys the second command a chance.
• Calibration improves: the share of your predictions that matched what Claude actually said.
• Nobody reports being blocked, interrupted at a bad moment, or having a prompt eaten. Those are correctness failures, not annoyances.

Where it could go wrong

• It requires honesty to work. Rate yourself 7 and look up the answer and you get nothing. There is no way to enforce this and no reason to try.
• Nobody wants to be tested. The core loop is mildly unpleasant by design. That is a real adoption risk, and softening it would delete the value.
• The AI parts cost money. Writing questions and grading answers run through your own Claude CLI or a gateway key: typically a few dollars a month, with the running total in vouch status. The unused check, the map and the cards need no AI at all.
• A real-time hook in someone's editor is dangerous ground. Sampled, cooled down, always skippable, structurally unable to fail the session. Every one of those has to keep being true.
• Early software. 0.1.1, and it says so. The commands here are built; the polish is not finished.

What’s next

• Sharpen question quality. It is the difference between a real probe and a quiz.
• Widen the curated service table, since cost at scale and failure modes are the parts nobody can look up quickly.
• No dates. It is a 0.1.