A methodology for automating code development under supervision

What it is

A methodology for automating code development in a supervised way. The work is split into phases and, between one phase and the next, there are automatic verification gates that check what has been done before anything is allowed to continue. If a gate fails, nothing moves forward — and the verdict is issued by a script, not by the agent itself.

There are two lanes, because not every task deserves the same paperwork. The light one is a single file per task, for everyday work. The heavy one is for work where being wrong is expensive: a design a human approves → a phased plan → execution that leaves evidence per phase → a review against a contract written before any of it started, so nobody can move the goalposts halfway through. A third lane hunts bugs, with two blind discovery passes and an audit of the fix itself.

What matters is where the truth about the work lives: in versioned files — design, plan, per-phase logs, review — inside the project itself, next to the commit they belong to, rather than in a chat window that disappears when the session closes. If I come back tomorrow, or somebody else does, or another agent does, the file is there and it says exactly what state each thing was left in.

Technical stack

Five skills written as Markdown contracts for Claude Code — inputs, outputs, permitted verdicts, limits — plus two Node scripts that run the gates: a validator that walks the checks for each stage (human approval present, index format, verdict literals, phase-to-log coverage, clean working tree) and a tool that folds the approved plan into one aggregate sha256. That hash is the key piece: it is checked from three sides — the current seal, the one each log claims it ran against, and a fresh recomputation — and from that the script can tell a plan edited after approval from one executed against a plan that was never validated.

The division of labour is deliberate: the model does what needs judgement, and the scripts do what a machine does better, which is compare. A model never eyeballs a hash. A versioned installer drops the skills into any project, and the whole trail stays as plain text files: nothing is needed beyond Node and Git.