When we open-sourced agent-doctrine, our workflow for taking coding agents from spec to tested PR, the prose rules turned out to be one of the most-asked-about parts. Here's what they are and why.

Simplified technical English

We borrowed from ASD-STE100, the controlled language used in aerospace maintenance manuals:

• one meaning per word
• one idea per sentence
• 8 to 20 words per sentence
• active voice

Answer, facts, stop

Every reply starts with the answer, then gives the supporting facts, then ends. No recap, no "let me know if you need anything else", no praise or apologies.

Every claim carries a fact

Each claim needs a number, file, name or decision. "Improved performance" fails. "Cut p95 from 800ms to 210ms in export.ts" passes.

Banned words and punctuation

No "robust", "seamless", "leverage" or "delve". No metaphors when a plain phrase works. No em dashes.

How it's enforced

The rules live in CLAUDE.md, which syncs to all seven supported hosts. A short per-prompt style reminder is also injected on every turn (there's an Amp plugin for it in amp/plugins/style-reminder). The reminder matters because long sessions drift, and a rule stated once at the top gets forgotten by turn 40.

Does it work?

Mostly. Reviews got faster because PR bodies say what changed and how it was tested, in that order. The rest of the doctrine (the pre-code gate, adversarial review) is in the repo if you want the full loop.