FINDS - TESTS — Slate 3, Test 4: No AI Slop (find #18) on a Sausalito blog draft
Date: 2026-11-20 · Tester: FINDS - TESTS subagent · Type: edit-quality test on a COPY. Nothing in ~/workspace/blog/ was modified; nothing was published.
Subject & provenance
- Repo:
petergyang/no-ai-slop, cloned from https://github.com/petergyang/no-ai-slop — exactly the repo named in find #18; identity confirmed by its README ("Remove 20+ patterns of AI slop from your writing without flattening your personal voice"). - Form: an agent skill, not a program:
skills/no-ai-slop/SKILL.md(rules + workflow) plusskills/no-ai-slop/eval.md(a self-check the editor must run on its own output). Packaged as a plugin for Claude Code / Codex / ChatGPT etc. - Modes: Edit (default — minimum effective edit, returns edited draft + a "What changed" section) and Detect (name patterns with quoted lines, no rewrite). A third README mode ("generate slop for fun") is satire, not tested. This test used Edit, the standard/default mode.
- Upkeep: 23 commits; tags v1.0.0–v1.0.6 (tags only — no GitHub Releases page verified); last commit 2026-09-01 (
000650b). Recent commits are rule refinements ("Remove metadiscourse…", "Add the portability test…", "Drop planning ritual…") — actively maintained, small and focused. License file present (MIT-style; see repoLICENSE).
Target selection
Inventory of ~/workspace/blog/ drafts:
| File | Words | Kind |
|---|---|---|
isometric-sausalito-00-intro.md |
976 | Series intro (ch. 0) |
isometric-sausalito-01-the-spark.md |
1,725 | Chapter 1 — substantive prose |
dendy-and-its-games.md |
~1,950 | Different series (Dendy) — out of scope |
one-bit-gaming-dendy-to-playdate.md |
~2,100 | Different series — out of scope |
yantr-self-hosted-app-store.md |
~1,070 | Different topic — out of scope |
publish/…, push-staging/… |
— | Staged copies of the Dendy posts, not drafts of this series |
Pick: Chapter 1, "The Spark: A Reel, Three Cities, and a $150 Cap" (1,725 words). It is the Sausalito series' first full prose chapter (the intro was dispreferred per the brief), dense with exactly what this test needs: personal voice, place-specific fact (Sausalito's area, water share, hillside mesh), named people/projects, and hard numbers. The draft's title-options block and production footer were treated as scaffolding and left unedited.
Method record
- Read
SKILL.mdandeval.mdin full; followed the skill's workflow literally: read the whole draft first, identified voice traits to preserve (bluntness, dry wit, numeric density, British-flavoured spellings, dash-driven cadence), then made the minimum effective edit, then ran theeval.mdchecks on the output (results inwhat-changed.md). - Mechanical pre-scan against the skill's own lists: banned words 0 hits (of 28), banned phrases 0 hits, often-empty adverbs 5 hits (
crucially×1,literally×1,just×1 geographic,actually×2 rhythmic/contrastive), em dashes 43 in the file / 32 in the prose body. - Edit pass: 19 operations (21 discrete changes; the phases sentence held two). Output preserved as
edited.md; the tool's "What changed" section aswhat-changed.md;edits.diffis the unified diff. - Judging pass: every change classified (a) improvement / (b) churn / (c) taste noise / (d) damage, plus a missed-items audit. Full table in
classification.md.
Results
- Words: 1,725 → 1,706 (−19, −1.1%). Prose em dashes: 32 → 8 (−75%), with 8 retained deliberately.
- Classification: (a) improvement 5 · (b) neutral churn 13 · (c) taste noise 2 · (d) damage 1.
- The genuine yield: removal of the draft's one true AI cadence — the binary contrast "The point of the gates wasn't caution for its own sake. It was that…" — plus a metadiscourse setup line and a list that now opens with a colon. All content-preserving.
- The churn engine is the em-dash rule. Most of the draft's dashes were appositive pairs doing ordinary punctuation work; swapping them for commas/parentheses changed nothing a reader would notice. Two swaps were actively worse than the original punctuation (taste noise), including a four-comma sentence replacing a cleanly dashed one.
- The damage is small but diagnostic: the binary-contrast rule, applied as written, deleted "That's not a confession;" from "That's not a confession; it turned out to be a design feature" — flattening a wry, characteristic pivot into a bland assertion. The rule cannot tell a formulaic "not X, it's Y" from a writer performing a denial on purpose.
- Why the yield is low: this draft was already nearly clean of everything the tool hunts. Zero banned words, zero banned phrases, sources named throughout (footer), numbers concrete everywhere. The tool is calibrated for raw AI output; this was a heavily human-worked draft. That is itself the finding for Ros's use case: his blog drafts go through heavy human/AI iteration before any "pre-publish pass" would run.
- What it missed (details in
classification.md): "270 times smaller" (broken math phrasing), zero fact-checking of the chapter's load-bearing figures, a referent ambiguity ("his companion 3D version"), "the plan" subject monotony, British/American spelling inconsistency ("amphitheatre", "artefact"), the comparison table's empty cells, and the bolded-four-decisions scaffold — an LLM-era structure its formatting rule can't see. - Internal rule conflict found:
eval.mddemands 1–2 em dashes per long draft; the skill's own principles demand minimum effective edit and cadence preservation. On this draft those cannot both pass. The edit pass prioritised the principles and logged the eval item as a deliberate strict-fail rather than butchering the prose to satisfy a quota. A user running this skill unattended would get whichever horn of the dilemma their agent happens to pick — the skill does not rank its own rules.
Credit where due
The skill is unusually restrained for its genre: "minimum effective edit," "leave strong human sentences alone," "would the writer recognize the edited draft as their own voice?" are its stated principles and its eval's final checks, and its Detect mode (name the pattern, quote the line, don't rewrite) is genuinely well-designed. Nothing in the run invented a fact, claim, or statistic — the "don't invent" rule held across all 19 operations.
Effort log
~40 minutes total: inventory + clone + repo read ~10 min; pre-scan + edit pass ~15 min; eval self-check, diff, classification and missed-items audit ~10 min; report ~5 min. No blockers. (The skill is executed by the testing agent itself, as with all prompt/skill tools — the edit pass and the judging pass were run as separate, sequential passes to keep them honest.)
Files
original.md— untouched copy of the draft as foundedited.md— the tool's edited draftwhat-changed.md— the tool's own "What changed" output + eval self-checkedits.diff— unified diff, original → editedclassification.md— change-by-change classification with quotes, counts, and the missed-items listrepo/— the cloned tool (petergyang/no-ai-slop @000650b)
Draft verdict + test-record block
Find #18 — No AI Slop (petergyang/no-ai-slop) - What ran: Edit mode (default) applied exactly per SKILL.md + eval.md self-check, 2026-11-20, to a copy of Sausalito blog Chapter 1, "The Spark" (1,725 words). Nothing in
~/workspace/blog/modified; nothing published. - Outcome: 1,725 → 1,706 words; 21 changes classified: 5 real improvements (24%), 13 neutral churn (62%), 2 taste noise (10%), 1 voice-damaging edit (5%). Missed 8 editor-level issues incl. a broken "270 times smaller" and all fact-checking. Found an internal rule conflict (em-dash quota vs voice preservation). Draft was already near-clean of the tool's target patterns (0 banned words/phrases), which caps the tool's yield on Ros's already-iterated drafts. - Output reference:~/workspace/finds-tests/no-ai-slop-2026-11-20/(report.md, edited.md, edits.diff, classification.md, what-changed.md) - Reuse proposal: REUSE LATER, with one scoped adoption. Adopt the Detect mode as a pre-publish scan for the Sausalito blog — it names patterns with quoted lines and touches nothing, which is the safe way to use it on voice-driven writing. Keep Edit mode supervised-only: useful as a first pass on raw AI-generated drafts (where its yield would be far higher), but never run wholesale on a finished chapter — expect ~60% punctuation churn and a small chance of it deleting a deliberate rhetorical move. Not a substitute for an editor: it does no fact-checking at all.