Does petergyang/no-ai-slop detect and rewrite AI-sounding text?

Ledger entry: no-ai-slop (Instagram carousel by SBX AI; free MIT skill). User ask: "try in the ai-lab as a playground."

Hypothesis

Applying the skill to AI-sounding text removes generic AI-writing patterns while preserving meaning and voice, and does not damage already-good human text.

Setup

  • Locate the skill repo from the carousel (https://www.instagram.com/sbx.intelligence/p/DdWTvhRCAf-/; find the GitHub repo from the post/caption). Read all of its files before installing anything. Record repo URL, commit hash, license. It must be an inspectable MIT repo; if not found, stop and report.
  • Install only inside the experiment directory (outputs/skill/), not into global ~/.claude/skills.

Fixed inputs (in source/inputs/, written before running anything)

  • 4 synthetic AI-slop drafts (English, ~150 words each; marketing, email, product blurb, LinkedIn post).
  • 2 human-written good samples (use public-domain prose or text written fresh; never Betec material).
  • 1 Hebrew sample (the carousel is Hebrew-audience; the ledger flags Hebrew quality as unmeasured).

Cases

  1. Detect: does the skill flag the known slop patterns in each draft? Count patterns flagged versus a hand-made list in source/expected-patterns.md (made before running).
  2. Rewrite: with-skill versus no-skill baseline (same model, same prompt "improve this text"). Blind A/B: a separate judge agent sees both outputs in random order without labels and scores clarity, specificity, voice preserved, meaning preserved (1-5).
  3. Do-no-harm: run on the 2 human samples. Pass = minimal edits (report edit distance) and the judge prefers the original or calls it a tie.
  4. Meaning check: list facts/numbers in each input; pass = all survive in the output.
  5. Hebrew: run the Hebrew sample; report honestly whether output is usable (the judge must read Hebrew; flag if it cannot).
  6. Repeatability: run case 2 three times per draft; report variance.

Deliverables

insights/RESULTS.md with the table of scores, which patterns it catches or misses, harm on good text, verdict on "worth keeping". demo/index.html: before/after of one real draft from outputs/.