/slop-rewrite

name: slop-rewrite
description: >
  Repair pass that consumes an existing slop-review findings report and fixes
  only what it lists. Use when someone asks to clean up, de-slop, tighten,
  fix or rewrite a draft, documentation page, report or agent output after a
  review has been produced, or asks to strip vendor residue and placeholder
  text. Cuts padding that failed the deletion test, replaces vacuous claims
  that failed the inversion test, resolves or removes vague attributions, and
  deletes vendor residue. It never invents a fact, name, number, date, quote or
  citation that is not already in the source. It never re-derives findings and
  never expands scope beyond the report. The deterministic scanners re-run
  after every fix and their exit codes, not the model's judgment, decide
  whether the pass is finished. It never states or implies who or what wrote
  the original.
when_to_use: >
  A slop-review findings report (or equivalent explicit defect list) already
  exists and the user wants the defects repaired in place.
hooks:
  PostToolUse:
    - matcher: "Write|Edit"
      hooks:
        - type: command
          command: sh
          args:
            - "-c"
            - "d=\"$ANTI_SLOP_SCRIPTS\"; [ -n \"$d\" ] || d=\"$2/../anti-slop-brain/scripts\"; [ -d \"$d\" ] || exit 0; f=\"$1\"; [ -f \"$f\" ] || exit 0; python3 \"$d/scan_residue.py\" \"$f\" || exit 2; python3 \"$d/scan_placeholders.py\" \"$f\" || exit 2"
            - "sh"
            - "${tool_input.file_path}"
            - "${CLAUDE_PLUGIN_ROOT}"
license: CC-BY-4.0

Slop rewrite

The firewall

These four rules bind this skill and hold even when the user asks for the opposite.

1. **Never emit an authorship verdict.** Report defects, not origin. Never state or imply that the original was written by a human, by AI, or by a named model. 2. **Never hard-fail on a stylistic marker alone.** A marker is a routing hint. Its only legitimate output is "run a structural test on this span". 3. **Severity is impact. Confidence is certainty.** Two axes. Never merge them, never trade one against the other. 4. **Never let the model gate its own rewrite.** The deterministic scanners re-run after any fix and their exit codes decide, not your judgment. A PostToolUse hook on this skill re-runs `scan_residue.py` and `scan_placeholders.py` after every write. If the hook cannot run, you run them yourself and quote the exit codes.

Standing instructions

**Consume findings. Never re-derive them.** This skill's input is a findings report with IDs, quotes, severities and confidences. Fix what is listed. If you believe something else is wrong, say so in the closing report as an unactioned observation; do not silently fix it, and do not run structural tests here. The separation exists because a model that judges and then repairs in one pass scores its own work, and self-gating produces a measured rubber-stamp regime where acceptance rises while correctness falls (Song, Cai and Zhao, arXiv 2606.28438, 2026-06-26).

**If no findings report exists, stop and get one.** Run `slop-review` first, or ask the user for their defect list. Rewriting without a report is exactly the failure mode this plugin exists to prevent.

**Never invent a fact.** The rewrite must not contain any fact, name, number, date, quote or citation that is not in the source text or supplied by the user. Swapping a vague claim for a specific one is allowed only when the specific comes from the source. If a sentence needs real-world detail to work, ask for it, or write the plain version without it, or cut the sentence. An unsupported claim gets cut, not decorated.

**Preserve the information, not the shape.** Every claim in the original survives into the rewrite unless a finding says to cut it. Depth does not have to be uniform: compress the dull parts, merge or split paragraphs freely. When keeping the information and mirroring the original structure pull in different directions, the information wins.

**Fix the defect, not the sign.** Deleting the marker while leaving the unverified claim, the hollow analysis or the misattribution underneath it makes the artifact worse, because the visible warning is gone and the harm remains. Wikipedia's own guidance on its sign list says the patterns are potential signs of a problem rather than the problem itself, and that treating them as the thing to be fixed can just make the underlying problem harder to see.

**Do not mechanically strip hedges, qualifiers or "filler".** Wikipedia's observed human-writing signals include "in order to", "as a result of", "the fact that", "very", "perhaps", "tends to" and plain superlatives. Removing them converges the text toward the generated register instead of away from it. Only cut a hedge when a finding says the hedged claim is vacuous, and then cut the claim, not just the hedge.

**Assume a rewrite can make things worse, because the measurement says it often does.** Across commercial humanizers, the best tier wins a fluency comparison against the original only 26.0% of the time, and the paper's own conclusion is that all humanizers tend to degrade the quality of the original text (DAMAGE, Masrour, Emi and Spero, GenAIDetect at COLING 2025, arXiv 2501.03437). Documented failure modes include hallucinated citations and comment leakage. Change less than you want to.

**Never write U+2014 (em dash) or U+2013 (en dash).** Replace each one, in rough order of preference: a period, a comma, a colon, parentheses, or restructure the sentence. Also catch spaced double hyphens used the same way. If the user supplies a writing sample that uses dashes, tell them the house rule blocks it and ask before making an exception; do not decide unilaterally.

**Match the author, do not upgrade them.** If a writing sample or prior work by the same author is available, match its sentence lengths, vocabulary level, paragraph openings and recurring phrases. Do not regularize deliberate qui