unslop

# Claude Code — paste both lines into any session, restart, type /unslop
/plugin marketplace add MohamedAbdallah-14/unslop
/plugin install unslop

Cursor, Windsurf, Cline, Gemini CLI, Codex, or the CLI work too. Install options →

---

See the difference

Same facts. Different voice. The hero above is the visual; the table below is the readable form.

I am writing to express my profound enthusiasm for the Marketing Coordinator position at Acme Corp. With over five years of experience navigating the dynamic landscape of digital marketing, I am confident that my comprehensive skill set and passion for innovation make me a robust candidate. I am particularly drawn to Acme's cutting-edge approach, and I would welcome the opportunity to delve into how my background aligns with your team's goals.

Five years in digital marketing, most of it in small teams where nobody else writes the landing pages, so I end up doing it. I saw your Marketing Coordinator posting yesterday. The part about running campaigns end-to-end instead of handing them off to an agency is what pulled me in — that's the work I actually like.

21 out of 21 blind LLM-judge runs preferred the unslop rewrite over the original. See Measured results.

---

Who actually uses this

Mostly engineers cleaning up agent output before it ships, and people writing things real humans will read — cover letters, essays, LinkedIn posts. If you're the second category, the non-developer guide skips the jargon. If you're the first: it's a CLI plus a hook plus a regex pass.

---

60-second start

[!TIP] Not a developer? Start with GETTING_STARTED.md. Plain English, three copy-pasted lines, real cover-letter examples.

Claude Code plugin (no clone, no install script)

Open any Claude Code session and paste:

/plugin marketplace add MohamedAbdallah-14/unslop
/plugin install unslop

Restart Claude. Type /unslop. Done.

You'll see a [unslop:BALANCED] badge appear in the statusline. Everything Claude writes from here on comes out in a human voice. Type stop unslop to turn it off, /unslop full to turn it up, /unslop-help to see everything.

Cursor, Windsurf, or Cline

git clone https://github.com/MohamedAbdallah-14/unslop.git

Open the folder in your IDE. The bundled rule files at .cursor/rules/unslop.mdc, .windsurf/rules/unslop.md, and .clinerules/unslop.md load automatically. Type /unslop in the chat panel.

Gemini CLI

git clone https://github.com/MohamedAbdallah-14/unslop.git && cd unslop
gemini extension install ./

Reads gemini-extension.json and loads GEMINI.md + the unslop skill into context.

OpenAI Codex

Clone the repo — the plugins/unslop/.codex-plugin/plugin.json bundle is auto-discovered by the Codex IDE extension.

Claude Code without the plugin system (manual hooks)

For forks, air-gapped setups, or when you want to see exactly which files get written:

git clone https://github.com/MohamedAbdallah-14/unslop.git
cd unslop
bash hooks/install.sh            # macOS / Linux
pwsh hooks/install.ps1           # Windows

What this does:

• Copies hook scripts to ~/.claude/hooks/ (flat, not a subdirectory)
• Registers SessionStart and UserPromptSubmit in ~/.claude/settings.json, merged safely via Node (never clobbers existing hooks)
• Wires the statusline so [unslop:FULL] shows when active

Idempotent. Re-run anytime to upgrade. The bash installer re-verifies settings.json state on each run; the PowerShell installer checks file presence only, so pass -Force on Windows if settings.json was hand-edited.

Standalone CLI (no IDE needed)

pip install unslop
unslop --deterministic path/to/file.md

Two modes: --deterministic (regex, no API) or default LLM mode (calls Claude). See unslop/README.md for the full CLI surface.

---

Measured results

Blind LLM-as-judge preference test. Claude Sonnet 4.5 compares each unslop rewrite against the original without knowing which is which. Seven fixtures, randomized A/B sides, 3 independent runs per fixture = 21 judgments.

Metric	Baseline	unslop (balanced, 3-run)
Blind humanness preference	—	100 % (21/21)
Humanized wins / ties / original wins	—	21 / 0 / 0
AI-ism reduction (rule-counted)	0 %	92.1 % (9-fixture suite, 2026-04-28)
Flat-paragraph count across suite	14	13
Preservation of code / URLs / headings	—	byte-identical

Every fixture wins 3/3 runs. Reproduce with python3 evals/perceived_humanness.py --runs 3 (needs ANTHROPIC_API_KEY). Archived at benchmarks/results/humanness/three-run-post-soul-fix-20260421.json.

[!NOTE] Humanness preference is measured by an LLM judge. Detector-score resistance is a different problem entirely. See How it stacks up and When it actually matters. Two different jobs; unslop is honest about both.

---

What you get

Five modes

subtle keeps the shape, scrubs the fingerprints. balanced is the default. full rewrites with opinion. voice-match mimics a sample. anti-detector does the burstiness and specificity moves that move GPTZero scores.

Preservation that actually holds

Code blocks, inline code, URLs, headings, YAML frontmatter, tables, blockquotes — byte-identical on the way out. Deterministic mode fails the run if anything drifts. LLM mode gets the same preservation list as an explicit instruction.

Also catches the newer tells: curly quotes, knowledge-cutoff disclaimers, vague attributions, title-case headings, repeated - **Label:** bullet stacks.

Six assistants, one source

Claude Code, Cursor, Windsurf, Cline, Gemini CLI, OpenAI Codex. The same skill loads in each one through whichever loading mechanism the platform supports. Single source of truth, synced by CI.

Real detector feedback

Opt-in CLI flag scores text against the TMR detector (99.28 % AUROC on RAID, 125 M RoBERTa), escalates through the mode ladder, prints what it tried.

Persistent voice-match

Save a numeric profile from a sample of your own writing — sentence-length variance, contraction rate, pronoun ratios. Reuse across sessions. No text samples are stored, so the tool can't learn to flatter you over time.

Pairs with Custom Styles

Anthropic Custom Styles sets the ceiling at generation; unslop catches residue afterwards. The ICLR 2026 Antislop paper formalizes that split.

Surprisal-variance reading

--surprisal-variance returns the DivEye signal — per-token log-probabilities from a local distilgpt2. Flat AI prose lands near 0.6–0.9; literary human prose often exceeds 1.5.

Reasoning-trace sanitizer

Strip <thinking> / <analysis> / <reasoning> / <scratchpad> wrappers and ## Plan sections from agent output before it ships. Opt-in. Sidecar file preserves the original trace.

Mode gating

--no-structural, --no-soul, and --no-audit turn off the newer aggressive passes for highly formal content (legal, compliance). Per-file opt-outs via HTML comments.

---

In the wild

The badge is the only UI. Everything else is silent — the hook fires on SessionStart, injects the activation rule into Claude's context, and tracks the mode in $CLAUDE_CONFIG_DIR/.unslop-active (fallback: ~/.claude/.unslop-active). No network calls. No telemetry.

---

Using it

Toggle modes mid-conversation

Phrase	Effect
/unslop	Turn on (balanced)
/unslop subtle	Light touch
/unslop balanced	Default
/unslop full	Strong rewrite
/unslop voice-match	Mimic a provided sample
/unslop anti-detector	Adversarial paraphrase
stop unslop  ·  normal mode	Off

Mode persists for the whole session.

Sub-skills

Skill	Trigger	What it does
unslop	/unslop	Active humanization for live responses
unslop-commit	/unslop-commit	Conventional Commits in human voice
unslop-review	/unslop-review	Direct, kind PR review comments
unslop-file	/unslop-file <file>	Rewrite a markdown file (preserves code, URLs, headings)
unslop-reasoning	/unslop-reasoning	Strip AI slop from chain-of-thought (over-hedging, loops)
unslop-help	/unslop-help	Reference card

Voice-match (persist your style)

unslop --save-voice-profile samples/my-writing.md   # one-time
unslop --voice-memory --mode full document.md       # uses saved profile
unslop --clear-voice-profile                        # delete

Storage: $UNSLOP_STYLE_MEMORY → $XDG_CONFIG_HOME/unslop/style-memory.json → ~/.config/unslop/style-memory.json. File is mode-0600; symlinks refused. Profile is numeric metrics only — no prose stored.

Strip reasoning traces (agent output)

Agent output often carries private reasoning wrappers (<thinking>, <think>, <analysis>, <reasoning>, <scratchpad>, <plan>) or markdown sections labelled ## Reasoning / ## Thought Process / ## Plan. Ship these into a final doc and you leak a process artifact the reader never wanted.

unslop --deterministic --strip-reasoning agent-output.md

On a file, stripped content is written to agent-output.reasoning.md next to the target. On stdin, the sidecar is discarded. The sidecar is gitignored by default because reasoning traces can contain process notes you did not mean to ship. Opt-in; default off.

Surprisal-variance reading

cat sample.md | unslop --surprisal-variance
# { "path": "<stdin>", "mean_log_prob": -2.83, "surprisal_stdev": 1.74,
#   "surprisal_cv": 0.61, "token_count": 412, "model": "distilgpt2" }

First call downloads distilgpt2 (~330 MB) via HuggingFace; subsequent calls are ~1 s on CPU. Override with --surprisal-model gpt2-medium for a stronger but slower reading. Source: Ganapathi et al., DivEye (arXiv 2509.18880, TMLR 2026). Requires pip install torch transformers. Set UNSLOP_SKIP_SURPRISAL=1 to disable.

Configure default mode

export UNSLOP_DEFAULT_MODE=full

Or ~/.config/unslop/config.json:

{ "defaultMode": "full" }

Resolution: env var > config file > balanced. Set to "off" to disable session-start activation entirely.

Live detector feedback loop

python3 -m unslop.scripts.fetch_detectors   # one-time: ~500MB of weights
unslop --detector-feedback file.md          # humanize, score, escalate, report

Escalation ladder: balanced → full → full + structural + soul. Reports the score at each step. It does not claim to lower scores — it just tells you where you are.

Use --detector-loop-aggressive for the longer five-step ladder:

unslop --detector-feedback --detector-loop-aggressive file.md

---

How it stacks up

Not every tool in this space solves the same problem. Here's the honest map.

	unslop	Anthropic Custom Styles	Undetectable.ai / StealthGPT / HIX	Plain LLM prompt
Works across 6 AI assistants	✅ one plugin	🟡 Claude.ai only	❌ web paste-box only	✅ anywhere
Runs offline (deterministic)	✅ regex mode	❌ cloud only	❌ cloud only	❌ needs API
Preserves code / URLs byte-exact	✅ validated	🟡 best-effort	❌ often breaks code	❌ drifts
Blind human-reads-more-human test	✅ 100 % (21/21)	🟡 not publicly measured	🟡 vendor-claimed, unverified	🟡 varies by prompt
Honest about detector limits	✅ documents < 0.5 pp	✅ doesn't claim defeat	❌ "99.8 % undetectable" claims	—
No paste-in-browser round-trip	✅ inline in your editor	✅ inline	❌ copy-paste workflow	✅ inline
Open source, MIT	✅	❌ proprietary	❌ proprietary	—
Free	✅	✅ on Claude.ai	❌ $10–30/mo	✅
Voice-match from your own writing	✅ numeric profile on disk	🟡 manual style prompt	❌	🟡 via prompt

unslop is a polish layer, not a detector-defeat tool. Commercial SaaS humanizers are a different category and mostly don't beat a second pass through a different model family plus five minutes of manual editing (Chicago Booth 2026 audit: median detector-accuracy drop ~6 points, not the claimed 40+).

---

Limitations

• Rewriting can degrade statistical watermarks like SynthID or green-list schemes. Side effect, not a feature. If provenance matters, watermark after unslop.
• Detector evasion isn't durable when the verifier has source-generation logs or retrieval access. Use anti-detector mode for false-positive defense, not academic misconduct.
• AI detectors over-flag non-native English. Liang et al. (arXiv 2306.04723) found GPTZero, OriginalityAI, and Crossplag flagged >50 % of TOEFL essays as AI-generated. Keep drafts and process notes when stakes are high.

---

FAQ

No. It changes how the reply sounds, not what it says. Ask for a cover letter, you still get a cover letter. Ask for feedback on your essay, you still get feedback. The facts, the advice, the answer — all there. Just without "Certainly! What a fantastic question!" around them.

Mostly no, honestly. My own testing against the TMR detector (99.28 % AUROC) shows deterministic surface rewriting moves scores by 0.0–0.2 pp. This matches the Adversarial Paraphrasing paper (NeurIPS 2025) predicting that exact outcome: modern detectors fingerprint on structural signals that synonym-swap rewriting cannot move.

What actually lowers detector scores, in order: (1) paraphrase through a different model family — GPT → Claude → Gemini, (2) burstiness, (3) specificity the model can't fake, (4) contractions and small fragments, (5) breaking predictable structure. Items 2–5 are what /unslop anti-detector does. Item 1 is a workflow you orchestrate.

Detectors also have a big false-positive problem. Liang et al. (Patterns 2023) found >50 % of TOEFL essays flagged as AI-generated. If a reader is running your work through a detector, document your process and keep drafts.

Turn it off for those. unslop trades precision for voice. For anything where a reader needs to follow the text exactly — a lease, a drug interaction warning, a deployment runbook — you want the robotic version. unslop is for text where the reader needs to like the text.

Deterministic mode preserves code blocks, URLs, headings, tables, blockquotes, and YAML frontmatter byte-identical. The risk isn't the tool breaking code; it's the rewriter smoothing a number you misremembered and making the wrong version sound confident. Re-verify facts after humanizing.

Not for the default plugin mode (it uses whatever assistant is already loaded — Claude Code, Cursor, etc.). Not for deterministic CLI mode (--deterministic, pure regex, no network).

You do need ANTHROPIC_API_KEY for: (a) default LLM CLI mode, (b) the evals/ humanness harness, (c) /unslop voice-match and full modes when running outside an assistant.

No telemetry, no analytics, no phone-home. The plugin's hook scripts run locally. The CLI calls whichever API you configured (Anthropic, or none with --deterministic). The voice-match cache is a numeric-only JSON file on disk at mode 0600, stored under $XDG_CONFIG_HOME/unslop/. No prose is persisted anywhere.

Three differences:

1. It's consistent. A prompt works for one message; the hook activates the rule every session and reinforces it at turns 8/16/24 to beat persona drift (RMTBench / HorizonBench 2026 measure >30 % degradation after 8–12 turns without reinforcement).
2. It's specific. The rule names dozens of patterns to drop (sycophancy openers, stock vocab, hedging stacks, transition tics, significance inflation) and gives structural targets (burstiness CV, sentence-length spread). "Write like a human" relies on the model's guess at what human means.
3. It's measured. The blind LLM-judge test and rule-based AI-ism counter run on every change. The 100 % preference / 92 % reduction numbers come from that harness, not vibes.

Each layer matches its host: Python for the file rewriter (CLI, HuggingFace integration, test ecosystem), JavaScript for Claude Code hooks (that's what the SessionStart / UserPromptSubmit APIs accept), markdown rules for every assistant that reads .cursorrules / CLAUDE.md / GEMINI.md / .windsurfrules. The sync.yml workflow keeps a single source of truth mirrored to every platform-specific location.

"Slop" is the term the LLM-evaluation community converged on for the residue of RLHF preference training — tricolons, sycophancy, stock vocab, tidy five-paragraph shapes. The verb "unslop" is the operation. Name was taken.

---

Docs

• GETTING_STARTED.md — plain-English on-ramp for non-developers (cover letters, essays, LinkedIn posts).
• unslop/README.md — the Python package and standalone CLI.
• docs/RESEARCH_AND_TECH.md — public reference: the research that informs shipping code, the tech stack, and the design choices that make unslop different.
• docs/research/ — 20 research categories, 120+ angle files, full implementation trace mapping each finding to the line of code it motivates.
• CHANGELOG.md — all releases.
• CONTRIBUTING.md — PR workflow, test gates, SSOT layout.
• SECURITY.md — vulnerability reporting.
• CODE_OF_CONDUCT.md — community guidelines.

---

The file-rewriter (unslop) placeholder-protects these in deterministic mode and fails the run if the validator detects they changed:

• Fenced code blocks (``` ... ```) — content and structure
• Indented code blocks (4-space)
• Inline code (`foo()`)
• URLs and markdown links
• Headings (whole line, text and level)
• YAML frontmatter at file start (---\n...\n---)
• Blockquotes (> lines and multi-line > blocks)
• Markdown tables (pipe tables)
• Quoted single-word examples — "delve" or "tapestry" stays put, because the word is being discussed, not used (use/mention distinction)

File paths, commands, technical terms, version numbers, and error messages stay exact when they live inside code blocks / inline code / URLs. Bare prose references to them are not separately protected; deterministic regexes only target prose patterns, so they usually pass through, but review the diff if your file mixes prose with identifiers.

LLM mode (default) receives the same preservation list as an explicit instruction. It cannot be byte-enforced the way deterministic mode is, so run the file through python3 -m scripts --deterministic afterwards if you need a hard guarantee.

det = handled by deterministic regex mode. llm = requires LLM mode (semantic rewrite, not regex).

Category	Examples	Mode
Sycophancy openers	"Great question!", "Certainly!", "I'd be happy to help"	det
Stock vocab	delve, tapestry, testament, navigate (figurative), embark, journey (figurative), realm, landscape, pivotal, paramount, seamless…	det
Hedging stacks	"It's important to note that", "It's worth mentioning", "Generally speaking", "In essence", "At its core"	det
Performative balance	A "however" appended to every claim	det
Transition tics	"Furthermore,", "Moreover,", "Additionally,", "In conclusion,", "To summarize," at start of a sentence	det
Em-dash pileups	More than two em-dashes per paragraph (bullet lists get a per-item budget)	det
Significance inflation	"marks a pivotal moment", "stands as a testament", "enduring legacy", "leaves an indelible mark"	det
Notability namedropping	"maintains an active social media presence", "a leading expert in", "renowned for his work"	det
Superficial -ing tails	", highlighting the importance", ", emphasizing its role" — filler participle phrases	det (full)
Copula avoidance	", being a reliable platform," → ", a reliable platform,"	det
Long-sentence run-ons	Sentences ≥20 words in flat-shape paragraphs split at safe boundaries (;, , but , , however, , em-dash)	det (Phase 1)
Parallel bullet soup	3+ bullets sharing first word merged into one sentence	det (Phase 1)
Missing contractions	"do not" → "don't", "it is" → "it's" where safe	det (Phase 5)
Filler phrases	"in order to" → "to", "due to the fact that" → "because"	det (full)
Negative parallelism	"No guesswork, no bloat, no surprises" tricolons	det (full)
False-range clichés	"from beginners to experts", "from humble beginnings to"	warning
Synonym cycling	utilize + leverage + employ in one paragraph	warning

Mode gating. subtle runs stock vocab only. balanced (default) runs everything tagged det plus Phase 1 structural and Phase 5 contractions. full adds filler phrases, negative parallelism, and superficial -ing. Use --no-structural or --no-soul to turn off the newer passes for highly formal content.

Don't humanize everything. Humanization trades precision for voice. For code, legal text, medical advice, security warnings, runbooks — you want robotic. Precision beats voice.

Humanize when a human reader will judge you on how it sounds:

• Resumes, cover letters, personal statements, bios
• College essays and applications
• LinkedIn posts, cold outreach, marketing copy
• Blog posts, newsletters, anything where the voice is the product

The two real levers

After reading the full compendium, it comes back to two moves. Everything else is decoration.

Subtract, don't add. AI tone isn't a thing you layer on top of pretraining. It's a residue from RLHF — the model was trained on preference data that rewards polite, hedged, tricolon-heavy prose. The fastest path to human-sounding text is removing those patterns, not sprinkling in "warmth". Adding warmth just adds sycophancy, and sycophancy is the loudest AI tell there is.

Engineer burstiness. Humans write sentences of wildly uneven length. Seven words. Then a twenty-three word sentence that develops one specific idea with a clause that earns its place. Then four. LLMs default to flat, uniform sentence length, and that's what detectors key on (Category 04). Vary it and half the AI tell disappears on its own.

AI detectors — the honest version

The academic consensus across Categories 05, 15, 16, and 18: the detection arms race is structurally unwinnable for detectors. Adversarial Paraphrasing (NeurIPS 2025) drops every tested detector's TPR by ~87 %. DIPPER did roughly the same thing in 2023. Detectors also have a huge false-positive problem on non-native English writers (Liang et al. Patterns 2023: >50 % of TOEFL essays flagged as AI). A flagged score means less than marketing pages suggest.

What I found running the TMR AI-text detector (99.28 % AUROC on RAID, 125 M-param RoBERTa) against the unslop pipeline on four AI-generated fixtures: deterministic surface rewriting — lexical + structural + contractions, every combination — moves the detector score by 0.0 to 0.2 percentage points. Scores stay pinned above p_ai = 0.98 regardless of what unslop strips. Adversarial Paraphrasing NeurIPS 2025 predicted exactly this: modern detectors fingerprint on structural signal that synonym-swap rewriting cannot move.

So unslop is a polish tool, not a detector-defeat tool. The blind LLM-judge test shows it decisively wins the "reads more human" comparison (100 %, 21/21). It doesn't fool GPTZero. Two different jobs.

What actually lowers detector scores, ordered by strength:

1. Paraphrase through a different model family. If GPT wrote it, have Claude rewrite. Or Gemini. Different stylometric fingerprints. The single strongest lever, and unslop cannot do it alone. TempParaphraser (EMNLP 2025) reports an 82.5 % average reduction in detector accuracy. When the --detector-feedback ladder exhausts, the CLI prints this recommendation explicitly.
2. Burstiness. Span sentence lengths roughly 4 to 35 words inside a paragraph. Phase 1 structural does this when material exists.
3. Specificity the model can't fake. Real dates, real project names, real numbers, first-person anecdotes. Training data doesn't contain your specifics.
4. Contractions and small fragments. "don't", "won't", the occasional start with "And" or "But". Phase 5 soul does the contraction half.
5. Break predictable structure. If every bullet has the same shape (verb + metric + with + tool), vary half of them.
6. One or two rough edges. A slightly awkward phrasing, a parenthetical trail, a non-linear logical jump — all read human.

Commercial humanizer SaaS (Undetectable.ai, StealthGPT, WriteHuman, HIX Bypass, Ryter Pro, Walter Writes AI, GPTHuman.ai — the ~150 products Category 18 audits) mostly don't beat a second pass through a different model plus five minutes of manual editing. Independent audits (DAMAGE COLING 2025; Epaphras & Mtenzi 2026; Turnitin's August 2025 anti-humanizer update) show wide gaps between their "99.8 % undetectable" claims and reality, and the gap shifts monthly. Chicago Booth's 2026 audit of twelve humanizer services found the median accuracy drop in downstream detectors was ~6 points, not the claimed 40+.

The right comparison isn't another SaaS. It's Anthropic Custom Styles (shipped November 2025 in Claude.ai) and OpenAI's style-steering prompt patterns — first-party style control from the model vendor, targeted at the same job. unslop is complementary: Custom Styles at generation time, the deterministic + LLM rewriting in this package after generation. The ICLR 2026 Antislop paper formalizes this split as "auto-antislop".

Resume playbook

The canonical case. Full stack in order:

1. Start with raw facts. Before touching an LLM, jot the bullets as notes. What you did, what changed, what the number was. No prose yet.
2. Use the LLM for structure, not voice. Ask it which accomplishment matters most, what's missing, how to order bullets. Don't let it write the final language.
3. Write the bullets yourself. Fast. One pass. Short. Specific numbers. Real tool names. The roughness of your first draft is the feature.
4. Polish grammar only. Tell the model: "fix typos and grammar, don't change word choice, don't smooth the voice, don't add adverbs." It will try to misbehave. Be strict.
5. Vary bullet shapes. Don't let every bullet read "Verb + metric + by using + tool". Some start with context, some with outcome, some with the action.
6. Top summary in your real voice. Not "Results-driven professional with a passion for". Something like: "Backend engineer. Ten years in payments. I like the unsexy systems work nobody volunteers for."
7. Human-read, not detector-read. If a friend says "yeah, that sounds like you", you're done. Detector scores are noisy and change weekly.
8. Optional paranoia pass. If the ATS is known to run detectors, paraphrase once through a different model family, then manually restore any bullet where the paraphrase killed a specific number or tool name. Never trust a paraphrase blind.

Persona drift over long sessions

RMTBench and HorizonBench (arXiv 2604.17283, April 2026) measure >30 % persona-consistency degradation after roughly 8–12 user turns in the same session. Two layers cover this:

• hooks/unslop-mode-tracker.js tracks a per-session turn counter (~/.claude/.unslop-turn-count) and re-emits an expanded reinforcement banner at turns 8, 16, 24, 32, and every 16 thereafter. No opt-in needed; the hook handles it. hooks/unslop-activate.js resets the counter on session start so nothing persists across shells.
• For voice-match, unslop/scripts/style_memory.py stores a numeric stylometric anchor on disk. Pure numbers, no free-text preferences. The MIT/Penn State CHI 2026 paper on "sycophancy memory" links free-text preference storage to amplified sycophancy over time. The cache makes that vector physically unavailable.

The warmth-reliability warning

[!WARNING] Training (or prompting) a model to sound warmer raises its error rate 8–13 % and amplifies sycophancy (Ibrahim/Hafner/Rocher 2025, Category 07). Fluent wrongness is worse than stiff accuracy, especially on a resume where a wrong date or an inflated metric can end the interview. After humanizing anything factual, re-verify every number, date, title, and tool name against the source.

/unslop anti-detector mode

An LLM-mode procedure. Covers items 2, 4, 5 from the detector list in one pass: burstiness targeting, contraction lift, structural variance. Item 1 (different-model paraphrase) the skill cannot execute alone — you have to request it. Use this mode when the reader might pipe the text into GPTZero or Turnitin. Skip for code, legal, or anything where precision beats voice.

My own testing: deterministic rewriting moves TMR scores by < 0.5 pp. Real detector resistance needs the different-model pass that only you can orchestrate. unslop's value in anti-detector mode is doing the local burstiness / contraction / specificity work correctly so the cross-model pass has less to fix.

---

Engineering & research

Every rule that ships in this repo ties back to a paper or a working open-source project. Not vibes. The full list lives in docs/RESEARCH_AND_TECH.md — 38 verified citations across 20 research categories, each one linked to the file and line of code it motivates.

Inspirations

Five projects and papers carry the most weight in shaping what unslop does and doesn't do.

Source	What it taught me
blader/humanizer	The original "scrub the AI residue" pattern. The deterministic regex layer in humanize.py started by porting its rule families and grew from there.
Antislop · Paech, ICLR 2026	Formalized the "auto-antislop" split between generation-time style control and post-generation residue cleanup. unslop sits on the second half of that split.
Adversarial Paraphrasing · Cheng et al., NeurIPS 2025	Predicted exactly the result I measured: surface rewriting moves modern AI-text detector scores by less than 1 pp. The reason detector.py recommends a cross-model pass when the local ladder exhausts.
DivEye · Basani, Chen et al., TMLR 2026	Surprisal-variance as a humanness proxy. surprisal.py runs a local distilgpt2 to compute the canonical 10-feature signal — flat AI prose lands near 0.6–0.9, literary human prose often exceeds 1.5.
Liang et al., Patterns 2023	Over 50% of TOEFL essays were flagged as AI by GPTZero. The ESL false-positive problem is the reason /unslop anti-detector exists — defensive use, not academic misconduct.

What I deliberately don't do

• Claim detector defeat. TMR detector AUROC 99.28; deterministic rewriting moves scores 0.0–0.2 pp on my fixtures. The README says so. The marketing doesn't.
• Store free-text style preferences. The voice-match cache is numeric only — sentence-length CV, contraction rate, pronoun ratios. The CHI 2026 sycophancy-memory paper (MIT/Penn State) links free-text preference storage to amplified sycophancy over time. unslop makes that vector physically unavailable.
• Add "warmth" to text. Ibrahim, Hafner & Rocher 2025 (arXiv 2507.21919) found warmer-sounding LLM output carries an 8–13% higher error rate. unslop subtracts AI residue rather than layering empathy on top.
• Send anything anywhere. No telemetry, no analytics, no cloud roundtrip. The plugin's hooks run locally; the CLI calls only the API key you configure.

Where to read more

• docs/RESEARCH_AND_TECH.md — public reference: every paper, every tech-stack decision, every differentiator with file:line evidence.
• docs/research/ — 20 numbered category folders (academic, industry, commercial, practical) covering 120+ angle files.
• docs/research/IMPLEMENTATION_TRACE.md — every research finding mapped to the line of code it motivates.

---

Architecture

The mermaid diagram below is the same picture in source form for grep-ability.

flowchart LR
  subgraph SSOT ["Source of truth"]
    S1[skills/unslop/SKILL.md]
    S2[rules/unslop-activate.md]
    S3[unslop/SKILL.md]
  end

  subgraph Sync ["sync.yml (CI on push to main)"]
    SY[Byte-identical propagation]
  end

  subgraph Mirrors ["Mirrored locations"]
    M1[.cursor/rules/]
    M2[.windsurf/rules/]
    M3[.clinerules/]
    M4[.claude-plugin/]
    M5[plugins/unslop/<br/>.codex-plugin/]
    M6[gemini-extension.json<br/>GEMINI.md]
  end

  subgraph Runtime ["Per-assistant runtime"]
    R1[Claude Code hooks<br/>SessionStart + UserPromptSubmit]
    R2[Cursor rules auto-load]
    R3[Windsurf rules auto-load]
    R4[Cline rules auto-load]
    R5[Gemini extension install]
    R6[Codex plugin discovery]
  end

  subgraph Python ["unslop Python package"]
    P1[humanize.py<br/>det + llm passes]
    P2[validate.py<br/>preservation checker]
    P3[structural.py<br/>Phase 1 burstiness]
    P4[soul.py<br/>Phase 5 contractions]
    P5[detector.py<br/>TMR / Desklib]
    P6[stylometry.py<br/>voice-match profile]
  end

  SSOT --> Sync --> Mirrors
  Mirrors --> R1
  Mirrors --> R2
  Mirrors --> R3
  Mirrors --> R4
  Mirrors --> R5
  Mirrors --> R6
  Python -. CLI / skill .- R1
  Python -. CLI .- R5
  Python -. CLI .- R6

  classDef ssot fill:#1F3D2A,stroke:#9BD4A9,color:#F7FBF8
  classDef mirror fill:#132019,stroke:#3A5443,color:#D6E7DB
  classDef run fill:#0F1A14,stroke:#7C9885,color:#D6E7DB
  classDef py fill:#132019,stroke:#D97757,color:#D6E7DB
  classDef sync fill:#3D2F1F,stroke:#E6C675,color:#F7FBF8
  class S1,S2,S3 ssot
  class M1,M2,M3,M4,M5,M6 mirror
  class R1,R2,R3,R4,R5,R6 run
  class P1,P2,P3,P4,P5,P6 py
  class SY sync

.
├── skills/                   # SSOT for the five agent-facing skills
│   ├── unslop/               — main mode
│   ├── unslop-commit/        — commit messages
│   ├── unslop-review/        — PR comments
│   ├── unslop-help/          — reference card
│   └── humanize/             — mirror of unslop file rewriter
├── unslop/                   # SSOT for the file-rewriter (Python + skill)
│   └── scripts/              — humanize, validate, structural (Ph1),
│                               soul (Ph5), detector (Ph3), stylometry (Ph4)
├── rules/                    # SSOT for the short always-on activation text
├── commands/                 # Claude Code slash commands (TOML)
├── hooks/                    # SessionStart + UserPromptSubmit + statusline + installers
├── .claude-plugin/           # Claude Code marketplace + plugin manifest
├── .cursor/                  # Cursor rules + skills (mirror)
├── .windsurf/                # Windsurf rules + skills (mirror)
├── .clinerules/              # Cline rules (mirror)
├── .agents/                  # Agents marketplace manifest
├── plugins/unslop/           # Codex plugin bundle
├── tests/                    # pytest unit tests
├── docs/research/            # optional research compendium (not part of the plugin bundle)
├── assets/                   # hero, statusline, section banners, social preview (PNG)
└── .github/workflows/        # CI + sync SSOT to mirrored locations

Source of truth: skills/unslop/SKILL.md, rules/unslop-activate.md, unslop/SKILL.md. The sync.yml workflow propagates these to every mirrored location on push to main.

---

Tests

python3 -m pytest tests/ -v               # Unit + integration (humanize + hook install)
python3 tests/verify_repo.py              # Repo integrity (manifests, mirrors, syntax, fixtures)
python3 benchmarks/run.py --strict        # Offline benchmark on AI-slop corpus, CI gates

• tests/unslop/ — 333 tests covering file-type detection; every deterministic rule family; structural rewriter (Phase 1); soul contractions (Phase 5); detector feedback loop (Phase 3); stylometry (Phase 4); humanness harness (Phase 6); preservation (code, URLs, headings, YAML, tables, blockquotes); end-to-end round trip. LLM tests are opt-in (UNSLOP_RUN_LLM_TESTS=1).
• tests/test_hooks.py — hook installer (fresh, idempotent, preserves custom statusline), unslop-activate.js banner, unslop-mode-tracker.js slash commands + natural language + stop phrases, statusline badge output, symlink refusal, CLAUDE_CONFIG_DIR honoring.
• tests/verify_repo.py — every SSOT mirror is byte-identical after sync, JSON manifests parse, all JS / Bash / PowerShell scripts are syntax-clean, fixture pairs round-trip, plugin + marketplace manifests are wired.
• benchmarks/run.py — runs humanize_deterministic over a corpus of AI-slop markdown and reports AI-ism reduction, per-paragraph flat count, sentences split, bullet groups merged, per-file structural integrity. --strict fails the build on any regression.
• benchmarks/check_regression.py — compares latest benchmark output against a pinned post-phase*.json baseline. Fails if AI-ism reduction drops > 2 pp, flat-paragraph total rises > 2, or preservation breaks. Runs in CI on every PR.
• benchmarks/detector_bench.py — opt-in AI-detector benchmark (TMR, Desklib). Downloads HF weights on first run. Scheduled weekly via .github/workflows/weekly-detector-bench.yml.
• evals/perceived_humanness.py — blind LLM-as-judge preference harness. Claude Sonnet 4.5 (default) compares unslop-rewritten vs original without side metadata.
• evals/ — additional LLM-driven A/B harness (llm_run.py + measure.py) for snapshotting baseline vs deterministic vs LLM unslop on a fixed prompt set.

---

Roadmap

Living list. PRs welcome — see CONTRIBUTING.md.

• v0.1 — Deterministic regex rewriter for sycophancy + stock vocab
• v0.2 — Multi-platform sync (Cursor, Windsurf, Cline, Gemini, Codex)
• v0.3 — Claude Code plugin via marketplace (2-command install)
• v0.4 — Phase 1 structural (burstiness), Phase 3 detector loop, Phase 5 soul contractions
• v0.5 — Stylometric voice-match profile, reasoning-trace sanitizer, DivEye surprisal-variance
• v0.6 — VS Code extension (native, not via Cline)
• v0.6 — Browser bookmarklet for web UIs (ChatGPT, Gemini web, Claude.ai)
• v0.7 — Multi-language support (Spanish, French, German slop patterns)
• v0.7 — Automatic different-model paraphrase pass for real detector resistance
• v1.0 — Stable plugin API, frozen SSOT schema

---

Contributing

PRs welcome. Read CONTRIBUTING.md for the test gates and the SSOT sync rules — edit the source-of-truth files, not the mirrors, or CI will revert your change. The CODE_OF_CONDUCT.md applies.

Found a security issue? See SECURITY.md.

---

Support the project

If unslop saved you from shipping a "comprehensive solution that leverages cutting-edge synergies", a star on the repo is the cheapest signal that tells me this is worth maintaining.

Other ways to help: file an issue with a before/after where unslop missed something or rewrote something it shouldn't have. Ship a PR for a new rule, platform adapter, or language. Run the evals on your own writing and tell me what scores you see. Cite the project if you write about AI humanization — I'd rather build on shared evidence than repeat marketing claims.

---

Read more

Long-form, behind the tool:

• Claude rewrote my resume and I couldn't send it, so I built unslop — the origin.
• The AI writing tic I couldn't stop seeing after building a humanizer — what unslop's pattern-detection trained the eye for.

Cross-posted on Medium and dev.to.

---

License

MIT. Use it, fork it, ship it.

Built by Mohamed Abdallah — senior Flutter engineer, OSS contributor on Flutter Favorite packages.

Built with careful human edits and a healthy suspicion of "delve".