ai-engineering: Skill for Claude Code

One pass: five lenses, one cross-read, a verdict inside the spec

What it produces

The specification's ## Council section: three machine-read headings, the two counts, a verdict, a recommendation and one first step. No transcript file and no page beside it — the section is the record and the record is the artifact.

The lenses, and none of them sees another

Cost, reversibility, the undecidable path, what is taken on trust, the example nobody wrote. Each is a question, not a personality. Each reads the spec and nothing else. Not the plan, not the chat, not another lens's answer.

That rule is not taste. Put one answer in a reader's context and a right answer turns wrong 66.5% of the time, against 10.3% for a plain re-ask (report 003). Nobody was named as its author. So the harm comes from the words being there, not from being told who wrote them.

Reading alone also buys ground. Human reviewers who do not confer raise 14 issues a session against 9 (report 003). And 70% (report 003) of what they find is seen by one reader only.

Every finding carries a command a reader can run to see the gap. One with no command is cut before its section is written, and it is listed under its own heading below so the count can be recomputed rather than believed.

The cross-read, inside the same pass

Each lens then sees the other four answers, relabelled and shuffled, and not its own. It answers two questions. Which of these findings is a false alarm, and what command shows it? And what did all of us miss?

It is never asked which answer is best. Ranking five good answers is the worst case in the measured work: one judge falls from 0.70 to 0.34 moving from pairs to a list (report 003). What the cross-read buys is aim: false alarms fall from 22% to 5.3% (report 003), and no true finding is lost.

A refutation carries a command. The command is run and its output is written down. One with no command is dropped, the same way a finding with none is.

A refuted finding is struck through and kept, with the refuting command beside it. It is never erased: a real gap killed by a good-looking answer must leave more than a number. Where one lens refutes and another agrees, the finding stays under the gaps and is not also listed under the refuted heading — one heading owns one finding, or the two counts double-count what the counter exists to keep honest.