Slop Rewrite

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 quirks and do not replace casual words with formal ones.

Per-finding repair rules

Closing report

After the rewrite, output this and nothing more elaborate.

## Repaired
- F-001 HIGH fabricated citation: removed the citation and the sentence.
- F-002 MEDIUM padding: cut 1 sentence, 24 words.

## Not repaired, and why
- F-005 LOW copula avoidance: confidence low, no style pass requested.

## Layer 0 re-run after the rewrite
| Scanner | Exit | Finding |
|---|---|---|
| scan_residue.py | 0 | clean |
| scan_placeholders.py | 0 | clean |
| lint_voice.py | 0 | clean |
| scan_refs.py | not run | no network |

## Unactioned observations
- <anything you noticed but did not fix, stated as an observation, with no
  severity and no rewrite attached>

## Net change
Words: 1,240 to 1,090. Claims removed: 2 (both unsupported). Claims added: 0.

"Claims added" must be zero. If it is not zero, you invented something. Go back and remove it.

Depth, loaded on demand

• Read ../../references/false-positives.md before actioning any LOW finding.
• Read ../../references/structural-tests.md if a finding's artifact is missing or unclear and you need to know what the reviewer should have produced, then send it back rather than filling the gap yourself.
• Read ../../references/markers-tier1.md only to understand what a finding refers to, never to generate new findings here.
• Read ../../references/code-markers.md and hand the work to slop-code if the artifact contains source code.