The month AI wrote 99.77% of my code

When I wrote every line myself, a vague requirement produced slow progress. I would stall, notice I did not understand the problem, and go and ask. The stalling was a feature — it was the system catching an under-specified request before it became code.

An agent does not stall. Give it a vague requirement and it returns something confident, well-formatted, consistently styled, and wrong, in about ninety seconds. The error has not gone away; it has just arrived faster and wearing better clothes. So the cost of being unclear went up, not down.

The hours I no longer spend typing go into four things: writing the specification, deciding the architecture, writing or verifying the tests, and reviewing the diff.

Specification is the biggest of the four. Before anything is generated I write down the exact behaviour, the field names, the status values, and what 'done' means — the acceptance criteria that decide whether the result ships. If a number, a colour, a timing or a status value is not written down somewhere, it does not get invented; it gets asked about.

Review is the second biggest, and it is the one that cannot be delegated. I read every diff. An agent that produces a thousand plausible lines an hour is only useful if someone is capable of telling which of them are wrong.

The interesting part of that statistic is not the 99.77%. It is the remainder. A small fraction of the lines still have to be mine — the ones where the trade-off is genuinely a judgement call, where a shortcut would be invisible now and expensive in six months, or where the agent has confidently solved the wrong problem.

Knowing which lines those are is the skill. Nothing about the last year has made it less valuable.