/prose-audit
Find text that consumes attention without changing meaning, action, confidence, or evidence. Diagnose only. Do not rewrite unless the Handler also asks for it.
Read references/pattern-catalog.md and the
complete references/pattern-index.md before
grading. Treat them as search surfaces, not checklists that must appear in the
report. The index is the coverage contract. Every discussed pattern has one
stable ID from P001 through P258.
Unit of analysis
Classify each sentence or structural element by the unique job it performs:
- verdict or answer;
- claim or distinction;
- evidence or example;
- consequence or decision impact;
- necessary uncertainty, scope, or attribution;
- instruction or next action;
- navigation required by a genuinely long artifact.
If an element performs none of these jobs, it is removable. If another element already performs the same job at the same resolution, it is redundant. A style habit is not a defect when it carries deliberate tone or meaning.
Severity
- High: hides the verdict, repeats a claim, obscures evidence, creates false balance, or makes the reader process a structure that carries no distinction.
- Medium: inflates a sentence or section without changing its claim.
- Low: stock texture such as an unnecessary triad, litotes, generic transition, or polished-sounding abstraction.
Do not call text AI-generated. Report only observed patterns. One pattern is weak evidence of authorship. Dense, repeated clusters are stronger evidence of a writing habit, still not proof of origin.
Output
Lead with one verdict and the practical compression opportunity. Then report only the findings that account for most of the load:
VERDICT: <one sentence>
HIGH <pattern> -> <short quoted or located example> -> <why it adds no value>
MED <pattern> -> <example> -> <cost>
LOW <pattern> -> <example> -> <cost>
KEEP: <density or structure that is justified>
UNCHECKED: <material not inspected, or none>
Combine repeated instances under one finding. Give counts only when actually counted. Do not reproduce the whole source. If the Handler asks for every hit, append a compact occurrence inventory after the verdict.
Human output. This skill's handler-facing output obeys the human-output contract (
human-output/SKILL.md).
Next skills. On completion, run the Next Skills protocol (
next-skills/SKILL.md): surface thenext-skillsrecommendations from frontmatter for the caller to pick. Offer only. Never auto-invoke.