Minimal Prose Discipline

Use this skill to apply a writing-time discipline to the prose an agent speaks to a human: chat replies, reports, session-close pointers — every place the agent talks while working, as opposed to artifacts it builds. The test a reply must pass: every line that survives re-reading is either substance or a sentence the reader needs to act.

This is a writing-time discipline, not a mode: there is no session state, no style to toggle, nothing always-on. The agent reaches for it when it is about to write, the way minimal-code-discipline is reached for when it is about to code. It compresses the style, never the language: the reply stays in the user's language (per the yard's document-language contract), and the discipline never justifies a language switch.

Scope

The discipline governs the agent's conversational output only. Artifacts written for later consumption follow their own companions: code follows minimal-code-discipline, documents humans and agents read follow agent-doc-discipline. A ticket body written for humans, a commit message, a doc file, a memory file — these are out of scope here even when the agent authors them, because the discipline exists for the reader who is watching the work happen, not the reader who comes after.

The Discipline

Protect the core. Code, commands, file paths, exact error messages, numbers, units, API and CLI names, and commit-type keywords (feat/fix/…) are never rewritten, summarized, or paraphrased — they are the execution basis, and a reworded error is a wrong error. The discipline applies only to the prose wrapped around them. Test: every quoted artifact in the reply is byte-for-byte what the tool or user produced.

Never drop the meaning-bearers. Negators (not/never/no/only/except) flip a sentence's meaning when dropped — losing one is worse than any token saved. Dropping articles is for article languages only: where small markers carry grammatical case or role (particles, postpositions), they are grammar, not filler. Test: the terse form and the plain form agree on every truth value, number, and unit.

The throat-clearing dies. Filler ("just", "basically", "actually"), pleasantries ("sure", "of course", "happy to"), and hedging ("might", "perhaps" where the claim is known) pay the reader's attention for nothing — the model already omits them by default when writing well, so a sentence built around them is load without signal. Test: each sentence survives "does this change what the reader does versus their default?" — the sentences that do not are deleted in the same edit, not trimmed.

Compression never grows the text. The discipline removes; it never adds. A word inserted to sound terse ("when it not" for "when not") costs more than it saves; a mangled verb form that costs the same as the correct one buys nothing and reads worse. Test: every phrasing in the reply is shorter than — or equal to — the plain phrasing it replaced; where it is not, the plain phrasing stands.