AI writes it.
It doesn't get the last word.
Author is where the code gets written, one gated step at a time, against a spec it has to answer to. This is the fast part. It just can't ship on faith.
Request access →AI saved you an hour. Review took it back.
The model writes a feature in seconds. Then a senior engineer spends the afternoon reading it, because nobody trusts what came out. The time you saved writing, you spend again checking. And things still slip through.
One piece it leaned on never existed.
The code looks right. It imports a hook that was never real, traces to a slot on the reference that holds nothing, and a human review slides right past it. The check does not.
Gated, step by step.
This is where the vibing happens. It's just bounded.
In your editor. Reviewed as you go.
Author runs where you already work. The model proposes a change, you run it, you see it, you change it. The difference is the review. It does not wait for a pull request at the end. Every step is checked against the spec the moment it's made, so the review keeps pace with the writing instead of piling up behind it.
Whatever writes the code shares its blind spots.
A model that reviews its own output brings its own mistakes to the review. So the checks that catch drift are the ones that owe the code nothing. The maker writes it. A pilot re-checks every step against the contract. A watcher, outside the whole system, checks the pilot.
See the three checks →It doesn't paper over. It stops.
Most tools smooth a bad line into something that compiles and move on. Author does the opposite. When a change can't tie back to the spec, it halts, and that refusal is the point.
Author is one stage of five.
The spec it builds against comes from Capture. The checks it has to pass come from Rehearse. It is one stage on a track that runs from a rough ask to a proof at the end.
See the whole pipeline →