unslop（unslop 搜索）

Unslop

Humanize AI-generated prose. Audit it first. Rewrite only when the user wants a rewrite.

When to Use

• User asks to "humanize", "de-slop", or "make it sound human"
• Editing AI-generated drafts, emails, articles, social posts
• Text contains AI patterns (throat-clearing, binary contrasts, em-dash abuse, emphasis crutches)
• User says text "sounds like AI" or "sounds robotic"
• Reviewing content before publishing
• User pastes text and asks to "clean it up" or "make it natural"

Arguments

Modes

This skill has two modes:

• rewrite (default) — diagnose, rewrite, then validate the rewrite
• audit-only — diagnose and assess the text without rewriting it

Trigger audit-only mode when the user says "audit only," "flag only," "scan this," "just detect," "don't rewrite," or passes --audit-only.

Voice Presets

Read the selected preset file from presets/ (e.g., presets/crisp-human.md) before writing. Each preset has specific voice rules, structural patterns, and a quality checklist.

Workflow

Pass 1: Diagnosis

Before rewriting anything, understand what's wrong. This prevents blind rewriting that loses meaning.

1. Read the relevant references based on what you see in the input:

- references/taboo-phrases.md — the complete catalog of banned phrases and newer structural families, with regex patterns for detection. This is the authoritative list; read it on first use and refer back for edge cases. - references/rubric.md — 8 scoring criteria (directness, rhythm, verbs, trust, authenticity, density, fact preservation, template avoidance), 5 points each. - references/fact-preservation.md — rules for what must survive transformation unchanged (numbers, names, dates, URLs, quotes, technical terms).

1. Extract constraints from the input — facts that must survive if you rewrite: python3 scripts/extract_constraints.py <<< "$INPUT" This outputs JSON with every number, date, name, URL, and quote that must appear in your output.
2. Scan for AI patterns: python3 scripts/banned_phrase_scan.py <<< "$INPUT" This returns violations grouped by category and severity (hard = always an AI tell, soft = context-dependent). Quoted examples, markdown blockquotes, and code snippets are ignored by default so you don't flag illustrative bad writing in docs. If you explicitly need to audit quoted examples too, run: python3 scripts/banned_phrase_scan.py --include-quoted <<< "$INPUT"
3. Read the selected preset from presets/ and note its voice rules.
4. Identify: audience, content type, tone target. A LinkedIn post needs different treatment than a technical doc.

Pass 2: Reconstruction

Skip this pass in --audit-only mode.

Rewrite the text. The references you read in Pass 1 are your guide — don't duplicate their rules here, just apply them.

Core principles (the why behind the rules):

• Em-dashes are the #1 AI punctuation tell. Default to zero. Use periods, commas, or parentheses instead. If one is absolutely necessary, max one per several paragraphs.
• AI text delays the point. Cut everything before the actual claim. "Here's the thing:" is throat-clearing. "Let that sink in." is an emphasis crutch. Just state the thing.
• AI inflates significance. "Stands as a testament to" means "is". "Pivotal moment" is almost never pivotal. Replace inflation with the specific fact.
• AI avoids commitment. "It's worth noting that" hedges. "Some experts argue" hides behind unnamed sources. Make claims directly or cite specifically.
• Facts are sacred. Every number, name, date, and URL from the original must appear in your output unchanged. Style is negotiable; accuracy is not.
• Shorter is almost always better. If cutting a sentence doesn't change the meaning, cut it. AI pads; humans compress.

Follow the preset voice characteristics for sentence length, paragraph structure, and tone. Refer to references/edit-library.md for 24 before/after transformation examples if you need guidance on specific pattern types.

For guidance on adding genuine human voice (not just removing AI tells), read references/personality-guide.md. Clean text that's still anonymous and voiceless scores a 3/5 on authenticity — aim for 4+.

Validation

Only run this section when you rewrote the text.

After rewriting, verify your work:

1. Fact preservation — confirm all constraints survived: python3 scripts/validate_preservation.py original.txt transformed.txt
2. Remaining AI patterns — check your output is clean: python3 scripts/banned_phrase_scan.py <<< "$OUTPUT"
3. Readability metrics — check rhythm and variance: python3 scripts/readability_metrics.py <<< "$OUTPUT"
4. Change percentage — flag if >40% changed (may indicate over-editing): python3 scripts/diff_check.py original.txt transformed.txt
5. Score against rubric — 8 criteria x 5 points = 40 max. Passing: 32/40 (80%). See references/rubric.md for detailed scoring.

Output Format

Adapt output to the context. For a quick fix, just return the cleaned text. For a thorough review, include validation:

Audit Only (--audit-only or user asks for flag-only scan):

## Issues Found

- [Quoted issue, category, why it reads as AI]

## Assessment

- [Which issues are clear problems]
- [Which issues are judgment calls or context-dependent]

Minimal (default for short text / quick fixes):

[The humanized text]

Detailed (for --strict mode or when user asks for analysis):

## Transformed Text

[The humanized version]

## Validation

- Constraints: [X]/[Y] preserved
- AI patterns: [N] remaining (was [M])
- Readability: Grade [X], sentence variance [Y]
- Change: [X]% from original
- Score: [X]/40

## Changes Made

- [List of major transformations applied]

Quick Examples

Input:

Here's the thing: building products is hard. Not because the technology is complex. Because people are complex. Let that sink in.

Output (crisp):

Building products is hard. Not the technology. The people.

Input:

In today's fast-paced business environment, it's becoming increasingly important for organizations to leverage their core competencies while navigating the complex landscape of digital transformation.

Output (crisp):

Companies need to use their strengths while going digital.

Reference Files

Located in this skill's directory. Read them as needed — don't front-load everything into context.

Scripts

All scripts accept stdin or file path arguments and output JSON. Run from the skill directory.

Maintenance Commands

Wiki Sync (/unslop --wiki-sync)

Syncs pattern rules with Wikipedia's Signs of AI writing page. Run periodically to pick up new patterns.

Steps:

1. Check for updates: python3 scripts/wiki_sync.py check (exit 0 = no updates)
2. Get structured diff: python3 scripts/wiki_sync.py diff (JSON output with change type, section, words)
3. For each new word/phrase: add to references/taboo-phrases.md in the matching section, and add to scripts/banned_phrase_scan.py BANNED_PHRASES dict with category, severity, and suggestion.
4. Verify: python3 scripts/banned_phrase_scan.py < /dev/null (confirm no syntax errors)

Only add phrases that are genuine AI writing tells for general prose. Skip Wikipedia-specific patterns (broken wikitext, DOI issues, etc.).

Key Principles

1. Diagnosis before writing — understand what's wrong before fixing
2. Facts are sacred — never sacrifice accuracy for style
3. Presets guide, don't constrain — adapt to content type
4. When in doubt, cut — shorter is almost always better
5. Quoted examples are exempt by default — don't flag illustrative bad writing unless the user explicitly wants that
6. Validation is mandatory — run the scripts, especially fact preservation