slop-gate

no-ai-slop is an editor you invoke. This is a gate that runs whether or not anyone remembers it. It scans outgoing text, refuses on a tell, and returns the hit list so the rewrite is targeted.

Two rules make it usable rather than annoying: hard tells fail on a single hit, and words that are decoration at volume but legitimate once are counted rather than banned. Text inside quotes or backticks is stripped before scanning, so a piece that quotes a tell as an exhibit does not fail on its own examples.

Why a gate and not a rule

A rules file is advisory, and advisory rules fail in a specific, repeatable way: they fire in the step the model thinks of as writing and not in the steps it thinks of as something else. A deck built by an agent can carry clean paragraphs under headline slide titles, because the titles were produced in a layout step where the writing rules never ran. Every surface nobody named explicitly sits on the unchecked side:

slide titles, kickers          table headers, row keys       commit messages
tile and card labels           diagram node and edge labels  code comments
chart titles and legends       button copy, subtitles        memory files
alt text, image captions       Slack, email, issue bodies    PR descriptions

The gate closes that by scanning the text on its way out, whatever produced it.

Run it

python3 scripts/scan.py DRAFT.md              # one file
python3 scripts/scan.py content/**/*.md       # many
cat draft.md | python3 scripts/scan.py -      # stdin

Exit code 0 is clear, 1 is tells found, so it drops into CI or a pre-commit hook unchanged. Every hit prints its label and the surrounding text.

Wire it as a hook (Claude Code, optional)

COG ships no hooks by default. If you want the check enforced rather than remembered, add this to .claude/settings.json yourself. --hook reads the hook payload on stdin and answers with a permission decision.

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Write|Edit|mcp__.*slack.*send.*|mcp__atlassian__(create|update)Confluence.*",
        "hooks": [{ "type": "command",
                    "command": "python3 \"$CLAUDE_PROJECT_DIR/.claude/skills/slop-gate/scripts/scan.py\" --hook" }]
      }
    ],
    "Stop": [
      { "hooks": [{ "type": "command",
                    "command": "python3 \"$CLAUDE_PROJECT_DIR/.claude/skills/slop-gate/scripts/scan.py\" --hook" }] }
    ]
  }
}

The Stop entry scans the agent's own reply, which is where most of these land in a working session. Expect it to fire on you within the hour; that is the point of it.

What fails on one hit

Tell	Specimen
Em dash	any —
Honesty framing	"honestly", "the honest part:", "if I'm being honest"
"not X, it's Y"	"It's not about speed, it's about trust"
"X, not Y" (trailing form)	"a test, not a permission", "machinery, not memory"
Rhetorical heading	"Why this matters", "The bottom line", "What comes next"
"The <Noun>" heading	"The Wedge", "The Shot", "The Gamechanger"
Verdict kicker	"That is the point.", "That is all a title is for."
Fake-profound closer	"let that sink in", "this changes everything"
Throat-clearing	"Here's the thing", "It's worth noting", "At the end of the day"
Sycophancy	"Great question", "You were right to push back", "That's on me"
Recap ending	a final paragraph opening "In conclusion" or "Ultimately"
Emoji heading	## 🚀 Results

What is counted, not banned

robust, seamless, leverage, utilize, delve, empower, streamline, genuinely, quietly, load-bearing, deep dive, clean, plain, simple, elegant, powerful, lightweight, tapestry, game-changer, paradigm shift, cutting-edge

Three or more in one piece fails, on the reading that density means decoration. One is a word choice. The test for any single use: remove the adjective, and if the reader would do nothing differently, it was decoration. "A robust retry" fails that test; "a retry that survives a mail-slot timeout" passes, because it names the behaviour. Keep the word when it is the term of art ("robust statistics").

Exhibits

A piece that must quote a tell (a style guide, a post about slop, a test fixture) carries slop-ok: <reason> anywhere in the file, usually as an HTML comment or a front-matter key. Quoted and backticked spans are already exempt, so the marker is only needed when a tell appears in the author's own voice on purpose.

Never add the marker to get past a gate you disagree with. Rewrite instead, or change the pattern list and say why.

What it cannot see

The gate is regular expressions, so it catches surface and near-surface tells. It is blind to the structural ones: a declared count that the source does not contain, a framework invented to organise two paragraphs, uniform sentence rhythm, a paragraph that restates the previous one. Those need a reader. Use no-ai-slop in detect mode, and prefer a reviewer from a different model family than the author, because a same-family reviewer shares the author's sense of what a finished answer looks like and waves through the same shapes.

For the tics that are yours rather than universal, measure them: voice-baseline.

Comparing models against the gate

scripts/model_compare.py replays your own Claude Code transcripts through scan.py and reports, per model, the share of chat replies and Markdown writes the gate would have refused and the hit rate per rule. It also reports workload per human prompt (tool calls, active minutes, tool error rate, subagent spawns) and punctuation per 10k words. When a hook refuses a reply, the transcript keeps the refused attempt as well as the rewrite, so the counts include what the model wrote before the gate saw it.

python3 scripts/model_compare.py --since 2026-09-15T00:00:00+00:00 --rules

Pick a window in which the rules file and hook did not change. A rule that tightened mid-window makes the earlier model look worse than it was.