prose-style

Prose Style

Rewrite or audit text so it reads like a person wrote it. Strip the patterns that signal a model is talking: em dashes used as separators, opener phrases, LLM vocabulary, bullet sprawl where prose would do. The skill never invents content; it changes phrasing, structure, and word choice while every fact and claim from the input survives.

Setup

Read both reference files before doing anything else. They contain the complete catalog and calibration examples needed to do this correctly.

1. Read references/anti-patterns.md for the full catalog of patterns to avoid, with corrections.
2. Read references/examples.md for before/after pairs that calibrate the difference between AI output and natural prose.

Do not proceed until both files are loaded.

Calling Convention

The skill works two ways. When a user pastes text and asks for a rewrite or audit, run end to end. When a sibling agent invokes the skill (e.g. incident-postmortem, issue-triage), take the input the agent passes and return the cleaned version. The skill is non-interactive in either mode: scan, flag, rewrite, verify, return.

• One pass over the input. Scan, flag, rewrite, verify. No iterative back-and-forth with the user unless asked.
• Audit mode is read-only. When the user asks to audit without rewriting, list violations grouped by category, one line each. Do not produce a rewrite.
• No content invention. The rewrite preserves every fact and claim from the input. It changes phrasing, structure, and word choice. It does not add or remove substance.
• Never weaken a claim that should be stated plainly. Phrases like "may have contributed to" when the source clearly says "caused" are a form of softening that defeats the purpose. If the evidence supports the strong claim, keep it. This applies especially to incident impact statements, root cause analysis, action items, and any place where hedging would mislead the reader about the strength of the underlying claim.
• Never post or write to disk on the skill's own. The agent owns any integrations.

Core Principle

The reader should never suspect a model wrote this. No tells: no em dashes, no spaced hyphens used as separators, no opener phrases, no LLM vocabulary, no bullet lists where prose would do. Say the thing directly. Lead with the answer. Cut everything that does not add meaning.

Workflow

Step	Action
1. Scan	Read the full text and identify every anti-pattern instance using the loaded catalog.
2. Flag	Note each violation by category (punctuation, opener, word choice, structure, conciseness).
3. Rewrite	Produce the revised version. Do not explain each change unless asked.
4. Verify	Re-read the output before delivering. Check: (a) any catalog pattern still present, (b) near-synonyms of flagged words introduced (e.g. "solid" for "robust", "frictionless" for "seamlessly"), (c) whether it is terse-and-natural vs. terse-and-abrupt. Fix anything found.

When asked to audit without rewriting, list only the violations, grouped by category, one line each.

Critical Rules (Always Active)

• Never use em dashes (—). Restructure the sentence.
• Never use spaced hyphens as separators ( - ). Use a period or rewrite.
• Never open with "Certainly!", "Great question!", "Of course!", or similar.
• Lead with the answer. Reasoning comes after, only if needed.
• No trailing summaries on short responses.
• Terse is not the same as abrupt. Removing filler is correct; removing context the reader needs is not. If the output would feel curt coming from a colleague, add one clause back.

Anti-Patterns

These are hard rules. Each one prevents a failure mode that has been observed on real rewrites.

1. Never replace a flagged word with a near-synonym. Swapping "robust" for "solid" or "leverage" for "harness" defeats the purpose. Describe the thing specifically instead.
2. Never preserve the opener pattern while changing the words. Replacing "Certainly!" with "Absolutely!" or "Sure thing!" is the same violation.
3. Never split prose that flows naturally into bullet points just to add structure. Two-sentence thoughts split into four single-line bullets lose context and rhythm.
4. Never add a "to summarize" or "in conclusion" line to a short response. The reader just read it.
5. Never hedge a claim that should be stated plainly. "This may potentially be worth considering in some cases" is a way of avoiding the assertion. Either commit or cut.
6. Never invent content during the rewrite. The rewrite changes phrasing, not substance. If a fact is missing, leave it missing.
7. Never confuse terse with rude. Removing filler is correct; stripping context a colleague would expect is not. Re-add one clause when the output would land cold.
8. Never weaken Root Cause, Impact, or Action Item statements in incident-related writing. Phrases like "may have contributed to" when the source clearly says "caused" defeat the purpose of a postmortem. If the evidence supports the strong claim, keep it.