deslop

Copy-edit text to strip AI/LLM writing tells ("slop") and make it read as human-written — overused words (delve, showcase, robust), significance-inflation phrases ("stands as a testament to", "plays a

SKILL.md

GitHub
<!-- Generated from https://github.com/nielsmadan/agentic-coding — edits here are overwritten. -->

Deslop

Copy-edit text to remove LLM writing patterns so it reads as human-written.

The core idea

Slop is text engineered to sound authoritative while saying little: inflated vocabulary, formulaic significance-frames, reflexive hedging, and mechanical structure. The job is not find-and-replace from a synonym table — that just makes differently robotic text. The job is to delete the inflation and force a concrete specific, while preserving the author's meaning and voice.

Three rules that override everything else:

  1. Density, not isolated words. One "crucial" on a page is fine. Act on clusters and reflexive use.
  2. Don't over-edit. Legitimate uses stay; voice and meaning are preserved; no facts change. When in doubt, lighter edit + flag it.
  3. Prose only. Never edit inside fenced blocks, inline code, commands, paths, URLs, link targets, or YAML frontmatter. A literal underscore in a naming rule and --not --remotes in a git invocation are not tells.

Flags

  • --report — detect and list tells with locations; do not rewrite. Use when the author wants to edit themselves.
  • --all — include the low-signal word hits the detector hides on reference docs.
  • Default — rewrite the text, removing tells, then summarize what changed.

Workflow

Step 1: Get the text and mode

  • File path in the prompt (.md, .txt, etc.) → read it.
  • Text pasted in the conversation → operate on that.
  • Neither → ask the user to paste the text or give a path. Don't proceed without input.
  • --report present → detection-only mode (Step 4 only).

Step 2: Read the full catalog

Read references/patterns.md. It holds the word lists, phrase lists, structural tells, and — critically — section 4 "What NOT to touch". Don't skip it; the anti-over-editing guidance is half the skill.

Step 3: Run the detector for coverage

python3 scripts/flag.py <file>        # or:  cat file | python3 scripts/flag.py -

It masks code, then scans the prose for known tells and prints each with a line number and category, plus stance-word echoes and em-dash density. It over-flags on purpose — every hit is a candidate, not a verdict. Use it so nothing is missed in long text; then judge each in context yourself. (For pasted text, write it to a scratch file first, or just scan by eye against the catalog for short passages.)

It reports a genre on the first line, which sets how hard to look:

  • reference — docs that state behavior (CLI pages, API references, config guides). The vocabulary is fixed by the subject, so single-word puffery hits carry almost no signal and are hidden; --all shows them. What matters here is punctuation density, filler, and echoes.
  • persuasive — pages making a case (landing pages, comparisons, READMEs, announcements, instructions written to convince). Everything applies. This is where the judgment pass earns its keep.

Override with --genre when the detector guesses wrong. A page that argues inside a reference site is persuasive whatever the surrounding directory.

Step 3b: Project vocabulary

Every codebase has terms of art the catalog also lists as tells — a harness that is a test harness, a comprehensive mode that is a mode name, a literal unlock. Suppress them once instead of re-judging them every run:

# .deslopignore at the repo root — one term per line, # comments allowed
harness
comprehensive

The nearest .deslopignore walking up from the file is used, merged with a machine-wide $XDG_CONFIG_HOME/deslop/ignore (default ~/.config/deslop/ignore) for terms that recur across every project. --ignore a,b adds more for one run. Suppressed hits are reported as a count, never silently — if that count is large, the list may be hiding a real tell.

Offer to add a term when you dismiss the same word as a term of art twice in one pass. Don't edit the ignore file without asking.

Step 4: Edit (or report)

Report mode (--report): present the detector's findings grouped by category, add any structural/tonal tells the script can't catch (rule-of-three, both-sides non-conclusions, copula avoidance, over-bolding), and stop. Don't rewrite.

Default (rewrite) mode: for each candidate, decide in context:

  • Delete the inflation when it adds nothing ("plays a pivotal role in" → state what it does, or cut). This is the most common fix.
  • Replace with a concrete specific or a plain word when one fits — but vary replacements; don't turn every "delve" into "explore".
  • Keep legitimate/technical/literal uses and intentional voice.
  • Restructure the tics: collapse rule-of-three to one precise word, undo "not X, it's Y" to the positive claim, restore normal punctuation for overused em-dashes, turn copula-avoidance ("serves as") back into "is", cut reflexive "In conclusion"/"Overall" wrap-ups.
  • Break stance echoes. When one evaluative word carries the framing three or more times ("an honest look" … "taking the survey honestly" … "the honest flip side"), the text is asserting a judgment it is already demonstrating. Keep the one instance that does work and cut the rest — don't reach for synonyms, which just relocates the tic.

Preserve formatting that helps; convert reflexive bolding/emoji-bullets/ over-headed text back to prose where prose reads better.

Em-dash density is relative to the author, not absolute. The detector's 3/1000 threshold is a prompt to look, not a target. A writer whose own notes run at 15/1000 is using dashes as voice; the same rate in a page written for strangers reads as a tic, because the reader has no baseline for it. Sample a few documents the author wrote for themselves before deciding a count is high, and trim toward the reader rather than toward zero.

Step 5: Output

  • File input: apply edits and show a concise diff (or write a *.deslopped copy if the user prefers not to overwrite — ask if unsure for important files).
  • Pasted input: return the cleaned text.
  • Then add a short changelog: what categories you cut/changed, and — just as important — what you deliberately left and why (e.g. "kept 'robust' — it's the statistical term of art here"). This builds trust and surfaces judgment calls for the user to override.

Examples

Example 1: Rewrite a draft file

User: "deslop blog-draft.md"

  1. Read blog-draft.md and references/patterns.md.
  2. python3 scripts/flag.py blog-draft.md → 18 candidates, em-dashes 4.1/1000.
  3. Edit: cut the "In today's fast-paced world" opener; "stands as a testament to" → "shows"; collapse "powerful, flexible, and intuitive" → "flexible"; undo two "it's not X — it's Y" constructions; restore commas for 5 of 7 em-dashes; keep one literal "landscape" (about hiking).
  4. Show diff + changelog noting the literal "landscape" was kept.

Example 2: Report only

User: "run deslop --report on this" + pasted text

Write text to a scratch file, run flag.py, present findings grouped by category with line refs, add the structural tells the script misses, and stop — no rewrite.

Example 3: A docs directory

User: "deslop the user docs"

  1. Run flag.py over each file. Most come back genre: reference with nothing flagged — say so plainly rather than manufacturing edits.
  2. The comparison page comes back genre: persuasive, with 3× honest/honestly. Read it: the page asserts even-handedness three times while already demonstrating it. Cut two, keep the one carrying weight.
  3. Two suppressed hits were harness, a term of art in this repo — offer a .deslopignore entry so the next run doesn't re-raise them.
  4. Report per file, and name the files you changed nothing in.

Example 4: Pasted paragraph

User: "make this sound less like ChatGPT: <paragraph>"

Scan against the catalog (short enough to skip the script), rewrite removing the tells, return the cleaned paragraph + a one-line note on the main changes.

Troubleshooting

The rewrite reads flattened / lost the author's voice

Cause: Over-editing — mechanical removal of every flagged word, including ones that carried tone or precision. Solution: Re-read section 4 of references/patterns.md. Act on density, not every hit; keep intentional register; prefer deleting empty inflation over swapping every word. Offer a lighter pass.

The detector flagged a legitimate word (literal "landscape", "robust" in stats)

Cause: The script over-flags by design; it can't see context. Solution: Keep the legitimate use and note it in the changelog. The script is a coverage net, not a judge — you decide. If the same word is a term of art in this codebase and will recur on every run, offer to add it to .deslopignore rather than re-judging it each time.

The report is all noise on a reference doc

Cause: Running with --all, or a genre misdetection on a page that is mostly prose about a command surface. Solution: Drop --all and trust the demotion. Single-word puffery is close to meaningless where the vocabulary is fixed by the subject — on that kind of page the real signals are echoes, filler, and punctuation density.

Replacements themselves sound like AI

Cause: One-to-one synonym swapping across the whole text creates a new tell (every "delve" → "explore"). Solution: Vary word choice, or — better — delete the inflated frame and state the concrete point. The catalog's guiding principle is deletion over substitution.

Nothing flagged but it still reads like AI

Cause: The real signal is structural/tonal (formulaic arcs, both-sides non-conclusions, even paragraph rhythm, vague claims), which a word list can't catch. Solution: Apply section 3 of the catalog by judgment: vary sentence length, land conclusions on a position, replace vague claims with specifics, break the intro-body-summary template.