Anti ai slop

Detects and removes AI-generated writing tells in plain text. Real specificity, preserved voice, no corporate sanitization.

Absolute bans (never ship)

These are non-negotiable in de-slop, humanize, and gate output unless the user explicitly overrides:

• No em dashes (—) or -- as dashes. Replace with commas, parentheses, periods, or colons. The #1 recognized AI tell.
• No meta-AI voice — "I hope this helps", "Certainly!", "Great question!", "I'd be happy to"
• No filler openers — "It's worth noting", "In today's fast-paced world", "Let's dive in"
• No formal transitions — Furthermore, Moreover, Additionally (use "And", "So", or nothing)
• No colon runways — "The result:", "Here's the key insight:" (state the fact directly)
• No Delvish clusters — delve + tapestry + nuanced + robust + landscape in the same piece
• No crucial-role formula — "X plays a crucial role in shaping Y"
• No isn't-it's pivots — "The hard part isn't X. It's Y." / "It's not X. It's Y."
• No lesson lists — "Two things that helped:" + takeaway bullets
• No kept/lost aphorisms — "Kept the badge, lost the pep talk"
• No fake-casual wit — "felt on brand", "No big scene, just…"

Full lists: registry/phrases.md, registry/structural.md.

Philosophy

1. The slop test: if a reader could say "AI wrote this" without hesitation, it failed.
2. Specificity beats polish — concrete nouns, verbs, numbers, constraints.
3. Preserve voice — remove tells; keep the author's register unless asked to change it.
4. Structural tells matter — uniform triplets, mirrored pairs, faux-profound closers, signposting scaffolds.
5. False-positive discipline — lists and frameworks are fine in moderation; Delvish words need clusters; em dashes have zero tolerance in output.

Setup

Skill root = directory containing this SKILL.md. Run all scripts from there:

node scripts/scan.mjs --json <target>
node scripts/gate.mjs --strict <target>
node scripts/check-fixtures.mjs

Before any command:

1. If the user invoked a sub-command, read reference/<command>.md. Non-optional.
2. Pipeline: linktwerk then de-slop → reference/pipeline-linktwerk.md.
3. Voice: default preserve → reference/voice-dial.md.
4. Pick a register (first match wins):
   • Czech email, outreach (Dobrý den, vykání) → registry/registers/email-cs.md
   • LinkedIn, job-hunt post, social → registry/registers/social.md
   • Email, essay, narrative (other) → registry/registers/prose.md
   • README, API docs, runbooks → registry/registers/docs.md
   • Taglines, hero copy, landing text → registry/registers/marketing.md
5. When a file path exists, run scan before audit, de-slop, or humanize:

   node scripts/scan.mjs --json <target>
   

6. Scan score is a floor, not a pass. Score 4 with LinkedIn/job-hunt/Czech email shape still requires LLM shape audit. See reference/audit.md.

Commands

Command	Purpose	Reference
scan [target]	Fast deterministic lint	reference/scan.md
audit [target]	Scan + rhetorical review, slop score 0–4	reference/audit.md
de-slop [target]	Aggressive rewrite; remove tells, keep meaning	reference/de-slop.md
humanize [target]	Light rewrite; max ~15% word change	reference/humanize.md
gate [target]	Pre-ship check; exit code pass/fail	reference/gate.md
loop [target]	De-slop + re-gate until clean (≤3×)	reference/loop.md
pipeline linktwerk	Cringe → human in one flow	reference/pipeline-linktwerk.md
ui-copy [file]	Scan extracted JSX/HTML text	reference/ui-copy.md

Scripts

Script	Purpose
scripts/scan.mjs	Lint + score
scripts/gate.mjs	Pass/fail gate (--strict = score ≥ 3, zero blockers)
scripts/check-fixtures.mjs	Regression test suite
scripts/extract-ui-copy.mjs	Pull copy from .tsx/.html

Slop score (0–4)

• 0 — obvious AI (5+ tells or dominant structural pattern)
• 1 — heavy (3–4 tells)
• 2 — noticeable (1–2 tells)
• 3 — mostly clean
• 4 — no tells; specific and human

Routing rules

1. First word matches a command → load its reference and proceed. Everything after the command name is the target.
2. No command, intent clear ("this README sounds robotic") → default to audit.
3. Target is UI (.tsx, .jsx, .html, .vue, .svelte) → extract copy via ui-copy; visual layout → impeccable.
4. User wants cringe LinkedIn only → linktwerk alone.
5. linktwerk then de-slop / linktwerk pipeline → pipeline-linktwerk.md (uses both skills in sequence).
6. Input is linktwerk output → de-slop + social register + loop.md.
7. Pasted text with no file → pipe to scan.mjs --stdin or gate.mjs --stdin --strict.

Post-output verification (de-slop / humanize / gate)

Before sending rewritten text to the user:

1. Re-read the draft aloud (mentally): does any sentence sound like a template?
2. Re-run scan on the draft (stdin or temp file).
3. Confirm zero em dashes, isn't-it's pivots, lesson-list headers, fake stat citations, punch-pause lines, hashtag walls.
4. If target was social/email: one concrete detail survives; no invented org names or survey percentages unless user provided them.
5. If scan score ≥ 3 but shape still feels AI → loop.md (max 3 iterations):

   node scripts/gate.mjs --json --strict draft.txt
   

Sibling skills

• impeccable — visual UI slop (gradients, card grids, layout).
• linktwerk — cringe generator; pair with pipeline linktwerk for human output.
• frontend-design — prevention at creation; this skill is review/fix at edit time.

Registry

• registry/phrases.md — banned and tired phrases
• registry/structural.md — pattern descriptions for LLM + script hints
• registry/registers/prose.md — essays, general email
• registry/registers/social.md — LinkedIn, job posts
• registry/registers/email-cs.md — Czech outreach
• examples.md — before/after pairs per register
• scripts/phrases.mjs — shared phrase lists for scanner (aligned with impeccable detect-text.mjs)

CI / regression tests

node scripts/check-fixtures.mjs
node scripts/scan.mjs --json --strict path/to/file.md

GitHub Actions runs check-fixtures.mjs on every push. Golden outputs live in tests/fixtures/golden/ and must pass strict gate.

Fixtures manifest: tests/fixtures.json.