no-slop

Write prose clean of AI texture, or audit a draft until it survives scrutiny.

npx skills add soheilmomeniii/no-slop

Or as a Claude Code plugin:

/plugin marketplace add soheilmomeniii/no-slop
/plugin install no-slop@somo-writing-skills

Every other de-slop skill removes AI tells. This one also checks whether the piece is telling the truth.

Two gates

Gate 1, source fidelity. Every quoted or attributed line is verified against primary sources and assigned one of four verdicts: verbatim, close-paraphrase, unverified, or misattributed. Silent paraphrase drift, fused quotes, and coinages credited to the wrong person all get caught. The standard is one line: if you cannot point to where a quote came from, it is not a quote yet.

Gate 2, taste. A mechanical scan for banned phrases, cadences, dashes, narrative cliches, and over-correction, then a nine-dimension scorecard scored by judgment. Fidelity beats polish on every tiebreak, because a shareable lie is still a lie.

Two rules a scanner cannot check

Never inject. A de-slop pass must not add a fact, number, name, stance, or first person the source never had. A fabricated specific is worse than the vague phrasing it replaced, so the skill leaves a marked placeholder and flags the gap rather than filling it.

Never overcorrect. Prose scrubbed to fragments with every hedge stripped is not clean, it carries a second fingerprint. The skill keeps roughness the author chose and the register the venue requires, and the scanner flags staccato runs so the failure is visible.

The scanner

python skills/no-slop/scripts/slop_check.py tics   <draft.md>
python skills/no-slop/scripts/slop_check.py quotes <draft.md> --corpus <sources-dir>

Pure Python, standard library only, no network, no subprocess, no shell. tics flags banned constructions and over-correction; quotes reports MATCH, PARTIAL, or NONE per quoted span against your sources and names the file each one matched. It reads curly single quotes as well as double quotes, verifies an ellipsis-elided quote segment by segment, in order, inside one source file, refuses to call it verified when the ellipsis omits a negation (text-perfect and meaning-inverted is the misquote a fidelity gate exists to catch), and never counts the manuscript as part of its own corpus. Exit codes are the contract: 0 clean, 1 findings, 2 error. Findings print to stdout, guidance to stderr, so the output pipes cleanly.

Exempt spans are blanked before matching, so a draft may quote bad writing on purpose without tripping the scan. The skill's own files pass their own scan, and a test enforces it.

Evals

Eleven scenarios in skills/no-slop/evals/, with fixtures. Four are adversarial by design: they punish a skill that scrubs too hard, that invents a specific to sound punchier, that strips hedges from an academic limitations section, or that flags an honest ellipsis-elided quote while missing a real cross-source fusion. On the v1.4.0 run recorded in CHANGELOG.md, ten independent agents passed ten of ten.

Layout

skills/no-slop/
├── SKILL.md                    the contract: modes, banned list, gates, report format
├── README.md                   the skill's own readme
├── references/
│   ├── fidelity-gate.md        Gate 1: verdicts, traps, correction log
│   └── taste-gate.md           Gate 2: scorecard, verdict criteria, carve-outs
├── scripts/slop_check.py       the scanner
├── tests/                      53 tests, including a self-scan gate
└── evals/                      11 scenarios plus fixtures
tools/validate_skills.py        spec + security validator, run in CI

Honest limits

The banned list is house style, enforced as taste. It is not an AI detector, and a hit is never proof that a text was machine-written. Published measurements find vocabulary lists barely separate machine prose from human prose, while rhythm uniformity separates them well, so treat every finding as a revision trigger and never as an accusation. For voice imitation use a style-converter skill; no-slop governs texture and truth, and composes with any voice on top.

License

MIT, see LICENSE. One section adapts Apache-2.0 material, credited in THIRD-PARTY-NOTICES.md.

{
  "name": "somo-writing-skills",
  "owner": {
    "name": "Soheil Momeni",
    "url": "https://github.com/soheilmomeniii"
  },
  "description": "Writing skills that keep prose honest and free of AI texture.",
  "plugins": [
    {
      "name": "no-slop",
      "source": "./",
      "description": "Write prose clean of AI texture, or audit a draft through a source-fidelity gate and a taste gate.",
      "license": "MIT",
      "category": "writing",
      "keywords": [
        "writing",
        "editing",
        "de-slop",
        "ai-detection",
        "fact-checking",
        "prose"
      ],
      "skills": [
        "./skills/no-slop"
      ],
      "strict": false
    }
  ]
}