unslop — Summary

unslop is a multi-platform AI output post-processor that strips AI-isms (sycophancy, stock vocabulary, hedging stacks, em-dash overuse, tricolon padding) from LLM responses while preserving technical accuracy and code blocks. It ships as a Claude Code plugin, Cursor/Windsurf rules, Gemini extension, Codex bundle, and a standalone Python CLI + evaluation suite. The skill persists across responses via a SessionStart hook that injects activation rules, with a UserPromptSubmit hook tracking mode changes.

The tool supports five intensity levels (subtle/balanced/full/voice-match/anti-detector) and is research-backed, citing studies on warmth-reliability tradeoffs, sycophancy rates, and verbal uncertainty calibration. A built-in eval suite (benchmarks/, evals/) measures "perceived humanness" against a 21/21 blind LLM-judge preference for rewrites. The tool explicitly bans specific vocabulary ("delve", "tapestry", "testament", "navigate" figuratively) and structural patterns (tidy five-paragraph shapes, em-dash overuse > 2 per paragraph).

Compared to seeds: no direct equivalent in the 11 seeds — unslop occupies a unique niche of AI output style correction rather than workflow orchestration or code quality. Closest to ccmemory in that it operates at the response level, but unslop modifies prose style, not knowledge retrieval.

unslop — Overview

Origin

Built by MohamedAbdallah-14. Released as MIT. The origin story: "Claude rewrote my resume and I couldn't send it. The polish was perfect; the voice wasn't mine. So I built this. It strips the AI residue and leaves the rest alone."

Philosophy

AI tone is a residue from post-training, not a layer you add with warmth. The right approach is subtraction (remove slop), not addition (add warmth). Adding warmth adds sycophancy — the loudest AI tell.

Five research-backed framing rules:

1. Subtract, don't add. AI tone is residue from post-training. Remove slop; never "warm up."
2. Style and stance are separate. Style = how it sounds; stance = how much it agrees. Move them independently. A humanized voice is not asking for agreement.
3. Warmth–reliability tradeoff is real. Ibrahim et al. (arXiv 2507.21919, 2025) found warmth-trained models had +11pp higher error rate on factual tasks. SycEval found sycophantic agreement in 58.19% of factual disputes across GPT-4o, Claude Sonnet, and Gemini.
4. Role-play frame, not personhood. Match the public style; don't invent biographical claims or suggest emotional investment.
5. Reason privately, humanize publicly. Do thinking in scratchpad; humanize only the final output. On reasoning-tier models (o1, o3, o4-mini, DeepSeek-R1), explicit CoT prompting adds no accuracy and increases variance 20–80%.

Explicitly Banned Terms

"delve", "tapestry", "testament", "navigate" (figurative), "embark", "journey" (figurative), "pivotal", "paramount", "nuanced" (when meaningless), "robust" (as filler), "seamless", "leverage" (when "use" works), "holistic", "comprehensive" (when "complete" works), "realm", "landscape" (figurative), "cutting-edge", "state-of-the-art" (as filler)

Banned Openers

"I'd be happy to help", "Certainly!", "Absolutely!", "Sure!", "What a fascinating...", "Great question!"

unslop — Architecture

Distribution

• Type: claude-plugin (primary) + multi-platform rule files
• License: MIT
• Install complexity: one-liner (Claude Code); clone-and-configure (others)

Install Commands

# Claude Code (plugin)
/plugin marketplace add MohamedAbdallah-14/unslop
/plugin install unslop

# Manual hooks (air-gapped or fork setups)
git clone https://github.com/MohamedAbdallah-14/unslop.git
bash hooks/install.sh    # macOS/Linux
pwsh hooks/install.ps1   # Windows

# Python CLI (standalone)
pip install unslop

Directory Layout

skills/
├── unslop/SKILL.md          # Main unslop skill (SSOT)
├── unslop-commit/SKILL.md   # Commit message de-slopping
├── unslop-file/SKILL.md     # File rewriting skill
├── unslop-help/SKILL.md     # Help skill
├── unslop-reasoning/SKILL.md
└── unslop-review/SKILL.md   # Review skill

commands/
├── unslop.toml              # Activate unslop mode
├── unslop-commit.toml
├── unslop-file.toml
├── unslop-help.toml
├── unslop-reasoning.toml
└── unslop-review.toml

hooks/
├── unslop-activate.js       # SessionStart hook: inject rules + write flag
├── unslop-mode-tracker.js   # UserPromptSubmit hook: watch for mode changes
├── unslop-statusline.sh/.ps1 # statusLine hook: show badge
└── unslop-config.js         # Shared module (flag file I/O)

rules/
└── unslop-activate.md       # Always-on activation rule

evals/
├── measure.py
├── perceived_humanness.py
└── prompts/

benchmarks/

unslop/SKILL.md              # File rewriter skill (top-level copy)

Source of Truth

Three SSOT files; all other copies are mirrors synced by .github/workflows/sync.yml:

• skills/unslop/SKILL.md — main skill
• rules/unslop-activate.md — always-on activation rule
• unslop/SKILL.md — file rewriter skill

Mirrored Platforms

.cursor/skills/, .windsurf/skills/, .clinerules/, .github/copilot-instructions.md, and others.

Required Runtime

• Node.js (for hook scripts)
• Python 3.x (for unslop CLI and evals)

unslop — Components

Skills

Commands (TOML slash commands)

Hooks

Intensity Levels

Eval Suite

• evals/measure.py — runs before/after eval
• evals/perceived_humanness.py — LLM-judge perceived humanness measurement
• Claims: 21/21 blind LLM-judge runs preferred unslop rewrite over original

unslop — Uniqueness & Positioning

Differs From Seeds

No direct equivalent in the 11 seeds. The closest in surface area is the anti-pattern lists in superpowers (which bans sycophancy in agent behavior), but superpowers targets agent decisions, not agent prose style. Unslop is the only tool in the corpus that operates purely on the register and tone of generated text, applying subtraction-first style correction backed by academic research on sycophancy rates and warmth-reliability tradeoffs.

Also distinct from all other tools in the corpus: this is not a workflow tool (it doesn't structure tasks), not a spec tool (it doesn't generate specs), not a verification tool (it doesn't check correctness) — it is a prose quality corrector. Unique niche in the AI-coding-agent ecosystem.

Observable Failure Modes

1. False positive vocabulary removal: Some "stock" terms may be technically appropriate in context (e.g., "robust" as a real technical description, not filler).
2. Over-application in technical contexts: The skill is designed for prose; code comments and docstrings have different norms.
3. Anti-detector mode risk: "Adversarial rewrite for AI-detector resistance" is a slower, more aggressive mode — may change meaning.
4. Multi-platform sync lag: The GitHub Actions sync workflow may lag; a platform-specific mirror may be stale.
5. Eval metric credibility: "21/21 blind LLM-judge runs" — LLM-as-judge studies have well-documented biases; this is a preference signal, not ground truth.

Distinctive Opinion

AI sycophancy is not a personality flaw — it's a training artifact that can be systematically removed with precise rules. "Warmth" added by the model correlates with factual error rates; removing AI-isms makes output more trustworthy, not just more readable.

Self-Referential Design

The README is written following unslop's own rules: terse, no sycophancy openers, before/after examples first. CLAUDE.md explicitly notes: "Preserve voice. The README's terse, no-AI-ism style is intentional — it is unslop eating its own dog food."