/deslop — AI-writing-tell scrubber

/deslop takes any public-facing text — a manuscript paragraph, grant prose, an email, a cover letter — and removes the fingerprints of AI writing so it reads like the user's own voice rather than model default. It runs two passes: a deterministic mechanical scrub, then a semantic rewrite into the user's cadence.

The detection taxonomy lives in references/ai-tells.md (tunable, versionable — the delve-list shifts as models change). The econometrics whitelist lives in references/domain-whitelist.md. The voice target lives in references/voice-samples.md, which holds two registers — Correspondence and Manuscript/grant — each falling back to the configured voice reference (personal_config.user.voice_style_ref) until populated. Read all three on every invocation — do not summarize from memory. The pass-2 rewrite uses ONE register, selected per the "Voice register" section below.

When to invoke

• "deslop this", "scrub the AI out of this", "make this sound less like AI"
• "does this read like AI?", "humanize this", "remove the AI tells"
• "de-slop my intro / email / cover letter / grant paragraph"
• Any time the user pastes text or names a file and wants the AI cadence gone.

Not for: verifying citations (→ /bibcheck), drafting new prose (→ /draft).

Inputs

The argument is either a file path or pasted text.

• File path (e.g. a .tex, .md, .txt): edit in place. Recoverable via

git / file history, so no review diff is needed (see Output behavior).

• Pasted text: return the cleaned text in chat. Do not write a file.

Flags:

• --voice=correspondence|manuscript — force the pass-2 voice register, skipping

auto-detection. Aliases: --voice=email → correspondence; --voice=paper and --voice=grant → manuscript. See "Voice register" below.

• --report — print the tell-scorecard only; do NOT apply any change. Use when

the user wants a preview.

• --mechanical-only — run pass 1 (deterministic scrub) and skip the LLM

rewrite. Faster, fully predictable, no voice judgment.

• --domain=marketing|econ|generic — selects the whitelist strictness

(references/domain-whitelist.md). Default: the user's field. generic relaxes the econometrics whitelist for non-technical prose (emails, letters).

Voice register — select before the pass-2 rewrite

The pass-2 rewrite targets ONE of two registers from references/voice-samples.md. The mechanical pass-1 scrub is register-independent (em-dash thinning, delve-list, artifact stripping, sentence-case headings, etc. apply to all text) — register selection affects only pass 2, plus the one punctuation rule noted below.

Select the register in this order:

1. `--voice=` flag wins. correspondence / email → Correspondence;

manuscript / paper / grant → Manuscript.

1. Auto-detect from the input:

• A .tex file, OR content containing \citep, \section,

\begin{abstract}, or \Cref, OR the user says "paper / manuscript / section / abstract / grant / proposal / R&R" → Manuscript register.

• An email, short message, cover letter, OR the user says

"email / message / note / reply / cover letter", OR short informal text with a greeting and/or sign-off → Correspondence register.

1. On ambiguity: infer the most likely register from content, APPLY it, and do

not stop to ask. STATE the choice in the scorecard (one line, e.g. Register: manuscript) so the user can re-run with --voice= if it was wrong.

Register-conditional punctuation rule (the one pass-1 exception):

• Correspondence: convert AI em-dashes "—" to a spaced hyphen " - " if that is

the user's correspondence convention (see references/voice-samples.md); never introduce an em-dash.

• Manuscript: em-dashes / en-dashes follow standard academic LaTeX usage. Do

NOT force the spaced hyphen. Still thin em-dash overuse by frequency, but a correctly-used em-dash in formal prose is fine.

Pass 1 — Mechanical scrub (deterministic)

Apply the regex + word-list detectors in references/ai-tells.md, in this order. Each detector is keyed on frequency, not presence — see the thresholds in the reference file. Run every delve-list candidate through domain-whitelist.md before flagging it.

Detector families (full patterns + thresholds in references/ai-tells.md):

1. Delve-list vocabulary — flag by density; suppress on a domain anchor.
2. Stock phrases — "it's important to note that", "stands as a testament",

"in the heart of", "rich tapestry of", etc.

1. Negative parallelism — not only … but, it's not …, it's ….
2. Vague attributions — "studies show", "experts argue" → named source or cut.
3. Connective-opener overuse — sentence-initial Additionally/Moreover/…
4. Trailing "-ing" significance clauses — "…, further enhancing its …".
5. AI-tool artifacts (HARD FAIL) — oaicite, contentReference,

turn0search, grok_card, stray utm_source=. Always strip on any match.

1. Punctuation / formatting — em-dash density, curly→straight quotes,

**Bold:** description lists, Title Case headings → sentence case, emoji in headings, --- before a heading. (Em-dash handling is register-conditional — see "Voice register": user's correspondence convention vs. standard academic usage in Manuscript.)

1. Rule of three — flag triplet density; prune the filler third in pass 2.

Pass 1 fixes are mechanical and safe (artifact removal, formatting, em-dash thinning, sentence-case headings). Anything requiring judgment about meaning is deferred to pass 2.

Pass 2 — Semantic rewrite to voice

Skip this pass under --mechanical-only or --report. Otherwise, first select the voice register (see "Voice register" above), then rewrite the flagged-but-not-mechanically-fixable prose into the user's voice for that register — the correspondence voice for emails/messages/cover letters, or the manuscript voice for papers/sections/abstracts/grants:

• Promotional / significance puffery → plain fact. "stands as a vibrant hub"

→ "is a town".

• Hedging filler → direct claim or deletion. Strip empty throat-clearing

("it's important to note that"); KEEP the author's genuine substantive hedges.

• Vague attribution → named source or cut. Never invent a citation.
• AI cadence → user's cadence. Kill elegant variation (forced synonym swaps),

kill the "Challenges and Future Prospects" conclusion frame.

• "is"-avoidance verb inflation → plain copula. "serves as / boasts /

represents" where it just means "is" → "is".

Ground the voice in the SELECTED register of references/voice-samples.md (each register falls back to the configured voice reference when its samples are empty; for manuscript input that reference is the same one /draft uses). Preserve every substantive claim, number, quotation, and title exactly — adjust phrasing and cadence only. For .tex, respect LaTeX: keep \emph over \textit, \Cref, natbib citations; do not mangle math or macros.

Domain whitelist

See references/domain-whitelist.md. The short version: a delve-list word is suppressed (not flagged) when it sits within a few tokens of a stats anchor (standard errors, regression, specification, identification, controls, p-value, …). "robust standard errors", "leverage points", "significant at the 5% level", "comprehensive controls" are technical vocabulary, not slop. The verb "leverage a dataset" IS a tell; the noun "leverage points" is not.

Output behavior — auto-apply (no review diff)

Per the user's explicit preference, do not run an interactive accept/reject loop per edit.

• File argument: edit the file in place with the Edit tool. The user

recovers prior versions via git / file history if needed.

• Pasted text: return the cleaned text directly in chat.
• `--report`: apply nothing; print only the scorecard.

After applying (or in --report mode), print a short scorecard — what changed, by category, with counts. Not a per-edit diff; a summary:

deslop scorecard — <file or "pasted text">  [domain=econ]
  Register: manuscript  (auto-detected; override with --voice=correspondence)
  HARD FAIL artifacts stripped ........ 2  (oaicite, contentReference)
  Delve-list words removed/replaced ... 5  (delve→examine, tapestry→set, …)  [3 suppressed by whitelist]
  Stock phrases rewritten ............. 3
  Em-dashes thinned ................... 4 → 1
  Negative parallelism rewritten ...... 1
  Vague attributions flagged .......... 2  (need a citation — see below)
  Title-case headings → sentence ...... 2
  Rule-of-three triplets pruned ....... 1
  Semantic rewrites (pass 2) .......... 6 sentences
Whitelist suppressions: "robust standard errors", "leverage points", "comprehensive controls"
Unresolved: 2 vague attributions need a real source — route to /cite or cut.

Anti-over-scrub guardrails

• Frequency is the signal, not presence. One em-dash, one triplet, one

"additionally" is fine. Tune thresholds; never zero out a construction.

• Preserve direct quotations and titles verbatim. Never scrub inside a quote.
• Preserve domain technical terms (the whitelist).
• Never invent facts to "improve" a sentence. If removing puffery leaves an

unknown fact, flag it — do not fabricate.

• Keep the author's real voice. Remove the machine's fingerprints, not the

writer's own style. The user's genuine hedges and rhythm stay.

Relation to /draft and /referee-response

• /draft writes new prose in the user's voice. /deslop cleans existing

text — they compose: draft → deslop, or deslop text written elsewhere.

• /referee-response writes R&R letters in-voice; run /deslop on a pasted

paragraph if it came from another tool and needs the AI cadence removed.

• /deslop never adds citations or verifies them — that is /bibcheck (audit)

and /cite (add).

Examples

Email (pasted, `--domain=generic`). Input: "I wanted to reach out to delve into a potential collaboration. Your groundbreaking work stands as a testament to your commitment to excellence — not only is it rigorous, but it also showcases a nuanced understanding of the field." → Output: "I'm writing about a possible collaboration. Your work is rigorous and shows a careful understanding of the field, which is why I'd like to talk." Scorecard: 1 stock phrase, 1 delve word, 1 negative-parallelism, 1 em-dash, "groundbreaking" puffery — all cleaned.

Manuscript paragraph (file, `--domain=econ`). Input contains "We leverage a rich tapestry of data. Our robust standard errors underscore the pivotal role of the treatment, highlighting its enduring significance." → The whitelist suppresses "robust" (anchored to "standard errors"). "leverage" (verb), "rich tapestry", "underscore", "pivotal role", and the trailing "-ing" clause are fixed: "We use a large panel. The treatment effect is precisely estimated (robust standard errors)." Scorecard notes the whitelist suppression so the user sees the term was deliberately kept.

Out of scope

• Citation verification — /bibcheck (audit existing .bib), /cite (add

one entry). /deslop only flags vague attributions; it never invents or verifies a citation.

• Writing new sections from scratch — /draft.
• Translating to a non-author voice (e.g. "make this sound journalistic") —

decline; /deslop targets the user's own voice only.

Credit

The AI-tell detection taxonomy in references/ai-tells.md (delve-list, stock-phrase set, negative-parallelism patterns, formatting tells, markup artifacts, vague-attribution markers) is derived from the Wikipedia essay "Signs of AI writing" — <https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing> — used under CC BY-SA. That derived taxonomy text carries the CC BY-SA attribution + share-alike obligation; the skill code / workflow itself is MIT (see the repo LICENSE). See ATTRIBUTION.md for the repo-level credit.