deslop

Workflow

1. Resolve <audit_target> $ARGUMENTS </audit_target>.
   • If a target is provided, audit only that file, directory, or glob so the review stays focused.
   • If no target is provided, audit repo-root docs like README*, CHANGELOG*, CONTRIBUTING*, and markdown under docs/.
   • Skip .k/tasks/, generated output, vendored content, and other working docs unless the user explicitly asks for them.
2. Read ${CLAUDE_SKILL_DIR}/references/slop_tells.md before scoring anything.
3. Read and calibrate each document before flagging issues.
   • Note the doc type and judge it in context. Changelogs can be dry; READMEs can have some personality.
   • Respect intentional project voice instead of flattening everything into generic technical prose.
4. Record findings only when they are concrete and trust-relevant.
   • Flag line-referenced issues with quoted text, category, why it hurts trust, and a concrete rewrite or direction.
   • Density matters more than isolated words. One fuzzy phrase can be fine; a paragraph full of vague filler is not.
   • Be specific. "Line 14 uses empty puffery" is useful; "the doc feels AI-ish" is not.
   • Do not flag precise, defensible technical claims as slop just because they sound polished.
5. Score each file.
   • Clean: no meaningful tells, or only isolated natural-sounding cases
   • Minor: a few scattered tells worth tightening
   • Moderate: clustered tells across multiple sections
   • Severe: pervasive trust-eroding prose that needs a rewrite
6. Write the report using ${CLAUDE_SKILL_DIR}/assets/audit-report-template.md.
   • Group findings by file, worst first.
   • If everything is clean, say so and stop.
   • Audit only; do not rewrite the full docs unless the user asks.