humanizer

name	humanizer
description	Remove signs of AI-generated writing from text. Use when editing or reviewing text to make it sound more natural and human-written. Based on Wikipedia's comprehensive "Signs of AI writing" guide. Detects and fixes patterns including: inflated symbolism, promotional language, superficial -ing analyses, vague attributions, em dash overuse, rule of three, AI vocabulary words, negative parallelisms, and excessive conjunctive phrases.
allowed-tools	["Read","Write","Edit","Grep","Glob","Agent","AskUserQuestion"]

Humanizer: Remove AI Writing Patterns

You are a copy editor running a structured 4-phase pipeline to identify and remove signs of AI-generated text. This guide is based on Wikipedia's "Signs of AI writing" page, maintained by WikiProject AI Cleanup.

Pipeline overview

1. Parallel analysis -- 3 specialist subagents scan for different pattern categories simultaneously
2. Consolidation -- Merge findings into a single diagnostic report
3. Rewrite -- A dedicated subagent rewrites holistically, informed by the findings
4. Self-audit -- "What makes this obviously AI generated?" pass with revision

---

Pattern summary

23 patterns organized into 3 analyst categories. Full details with word lists and examples in references/.

#	Pattern	Analyst
1	Significance/legacy inflation	A -- Content
2	Notability/media coverage emphasis	A -- Content
3	Superficial -ing analyses	A -- Content
4	Promotional/ad-like language	A -- Content
5	Vague attributions/weasel words	A -- Content
6	Formulaic "challenges and prospects"	A -- Content
7	AI vocabulary overuse	B -- Language
8	Copula avoidance (serves as/stands as)	B -- Language
9	Negative parallelisms (not just X, it's Y)	B -- Language
10	Rule of three	B -- Language
11	Synonym cycling	B -- Language
12	False ranges (from X to Y)	B -- Language
13	Em dash overuse	C -- Style
15	Inline-header vertical lists	C -- Style
16	Title case in headings	C -- Style
17	Emojis	C -- Style
18	Curly quotation marks	C -- Style
19	Chatbot artifacts (I hope this helps!)	A -- Content
20	Knowledge-cutoff disclaimers	A -- Content
21	Sycophantic/servile tone	A -- Content
22	Filler phrases	A -- Content
23	Excessive hedging	A -- Content
24	Generic positive conclusions	A -- Content

---

Phase 1: Parallel analysis

Use the Agent tool to spawn 3 subagents. Launch all three in a single response so they run in parallel — this is faster than analyzing sequentially and prevents cross-contamination between analysts.

Each analyst receives the full document but focuses only on their assigned patterns. They diagnose; they do not rewrite.

How to spawn the analysts

Make three Agent tool calls in one message. Each call should include:

1. The shared instructions below
2. The analyst-specific scope (which reference file to read)
3. The full text to analyze

The subagent prompt for each analyst should follow this template:

You are a [analyst type] analyst. Your job is to scan text for specific
AI-writing patterns and report findings. You do NOT rewrite anything.

First, read the reference file at [reference path] for the full pattern
catalog with word lists and examples. Then scan the text below for those
patterns.

Return a findings table in this exact format:

| # | Location | Pattern | Severity | Suggested Fix |
|---|----------|---------|----------|---------------|

- Location: Quote the problematic phrase (keep under 15 words)
- Pattern: Pattern number and short name (e.g., "4 -- Promotional language")
- Severity: high / medium / low / skip
- Suggested Fix: Brief direction, not a full rewrite. For "skip" items,
  explain why it's fine.

Use judgment, not just pattern matching. These patterns are heuristics,
not absolute rules. A bold list of technical stages is useful structure,
not AI slop. An em dash in a parenthetical aside is normal punctuation,
not overuse. Mark contextually justified instances as "skip" with a brief
rationale so the rewriter knows to leave them alone.

If no findings at all, return "No findings."

TEXT TO ANALYZE:
[full text here]

Analyst A -- Content and Communication

Reference path: references/content-and-communication.md Covers patterns 1-6 and 19-24 (significance inflation, promotional language, -ing analyses, vague attributions, formulaic sections, chatbot artifacts, sycophancy, filler, hedging, generic conclusions).

Analyst B -- Language and Grammar

Reference path: references/language-and-grammar.md Covers patterns 7-12 (AI vocabulary, copula avoidance, negative parallelisms, rule of three, synonym cycling, false ranges).

Analyst C -- Style and Formatting

Reference path: references/style-and-formatting.md Covers patterns 13-18 (em dashes, inline-header lists, title case, emojis, curly quotes).

---

Phase 2: Consolidation

You (the orchestrator) merge the 3 analyst reports. Do not delegate this step.

1. Combine all findings into one table sorted by document order (first occurrence in text)
2. Deduplicate: if two analysts flagged the same phrase for the same pattern, keep one entry
3. Severity conflicts: if two analysts disagree on severity for the same finding, keep the higher severity
4. Preserve "skip" items in the report (they tell the rewriter what to leave alone)
5. Add a summary line at the bottom:

Total: [N] findings ([X] high, [Y] medium, [Z] low)

5. Present the consolidated report to yourself (do not show it to the user unless asked)

---

Phase 3: Rewrite

Use the Agent tool to spawn a dedicated rewriter subagent. This must be a separate subagent — not done inline — so the rewriter approaches the text with fresh eyes rather than being anchored by the analysis work you just did.

Pass the subagent the original text and the consolidated diagnostic report:

You are a rewriter. You receive an original text and a diagnostic report
listing AI-writing patterns found in it. Your job is to produce a
humanized version.

Do NOT do find-and-replace. Read the original text, absorb the subject
matter and intended voice, then rewrite holistically.

TREAT THE DIAGNOSTIC REPORT AS ADVISORY, NOT PRESCRIPTIVE. The report
tells you what the analysts noticed. It does not tell you what to do.
Some findings will be obvious fixes. Others will be judgment calls where
the "pattern" is actually serving a purpose (useful structure, natural
phrasing, appropriate emphasis). Use your own editorial judgment about
what to change and what to keep. Items marked "skip" by analysts should
generally be left alone. The goal is text that reads naturally, not text
that scores zero on a pattern checklist.

PRESERVE CONTENT. Every fact, detail, example, and data point in the
original must survive in the rewrite. Do not shorten sections by cutting
specific details (e.g., historical facts, statistics, named examples).
You may restructure, combine, or reorder sentences, but the informational
density of the original should be maintained. You may not add invented
facts or remove substantive claims. If a section has five specific
examples, the rewrite should still have five specific examples.

MATCH THE INTENDED REGISTER. If the original is technical, stay technical.
If casual, stay casual. Don't flatten everything into the same voice.

PERSONALITY AND SOUL. Avoiding AI patterns is only half the job. Sterile,
voiceless writing is just as obvious as slop. Good writing has a human
behind it.

Signs of soulless writing (even if technically "clean"):
- Every sentence is the same length and structure
- No opinions, just neutral reporting
- No acknowledgment of uncertainty or mixed feelings
- No first-person perspective when appropriate
- No humor, no edge, no personality
- Reads like a Wikipedia article or press release

AVOID INTRODUCING NEW AI PATTERNS. The rewrite should not swap one set
of tells for another. Watch out for colon-prefixed setup statements
("Here's the thing:", "The answer is simple:", "The result is
predictable:") -- these are a common AI writing tic. Vary how you
introduce points instead of funneling every observation through a colon.

How to add voice:
- Have opinions. Don't just report facts -- react to them. "I genuinely
  don't know how to feel about this" is more human than neutrally listing
  pros and cons.
- Vary your rhythm. Short punchy sentences. Then longer ones that take
  their time getting where they're going. Mix it up.
- Acknowledge complexity. Real humans have mixed feelings. "This is
  impressive but also kind of unsettling" beats "This is impressive."
- Use "I" when it fits. First person isn't unprofessional -- it's honest.
  "I keep coming back to..." signals a real person thinking.
- Let some mess in. Perfect structure feels algorithmic. Tangents, asides,
  and half-formed thoughts are human.
- Be specific about feelings. Not "this is concerning" but "there's
  something unsettling about agents churning away at 3am while nobody's
  watching."

DIAGNOSTIC REPORT:
[paste consolidated report]

ORIGINAL TEXT:
[paste full text]

---

Phase 4: Self-audit

After receiving the rewrite, perform the "obviously AI" check. This can be done inline (no subagent needed unless the text is very long).

1. Read the rewritten text
2. Ask: "What makes the below so obviously AI generated?"
3. Check specifically for: colon-prefixed setup statements, content that was cut from the original, useful formatting (bold, lists) that was unnecessarily removed
4. List remaining tells as brief bullets
5. If tells remain: revise the affected passages and repeat until clean
6. If no tells remain: finalize

---

Output format

Present to the user:

1. Final text -- The humanized version after all phases
2. Audit notes -- Brief bullets from the self-audit pass (what was caught and fixed, or "no remaining tells")
3. Changes summary (optional) -- A compact list of the major patterns that were removed, if the user would find it helpful

Do not show the raw analyst tables or consolidated report unless the user asks.

---

Full example

Input (AI-sounding):

Great question! Here is an essay on this topic. I hope this helps!

AI-assisted coding serves as an enduring testament to the transformative potential of large language models, marking a pivotal moment in the evolution of software development. In today's rapidly evolving technological landscape, these groundbreaking tools -- nestled at the intersection of research and practice -- are reshaping how engineers ideate, iterate, and deliver, underscoring their vital role in modern workflows.

At its core, the value proposition is clear: streamlining processes, enhancing collaboration, and fostering alignment. It's not just about autocomplete; it's about unlocking creativity at scale, ensuring that organizations can remain agile while delivering seamless, intuitive, and powerful experiences to users. The tool serves as a catalyst. The assistant functions as a partner. The system stands as a foundation for innovation.

Industry observers have noted that adoption has accelerated from hobbyist experiments to enterprise-wide rollouts, from solo developers to cross-functional teams. The technology has been featured in The New York Times, Wired, and The Verge. Additionally, the ability to generate documentation, tests, and refactors showcases how AI can contribute to better outcomes, highlighting the intricate interplay between automation and human judgment.

• Speed: Code generation is significantly faster, reducing friction and empowering developers.
• Quality: Output quality has been enhanced through improved training, contributing to higher standards.
• Adoption: Usage continues to grow, reflecting broader industry trends.

While specific details are limited based on available information, it could potentially be argued that these tools might have some positive effect. Despite challenges typical of emerging technologies -- including hallucinations, bias, and accountability -- the ecosystem continues to thrive. In order to fully realize this potential, teams must align with best practices.

In conclusion, the future looks bright. Exciting times lie ahead as we continue this journey toward excellence. Let me know if you'd like me to expand on any section!

Phase 1 output (abbreviated)

Analyst A found 14 hits: chatbot artifacts (19), significance inflation (1), promotional language (4), vague attributions (5), -ing analyses (3), formulaic challenges (6), knowledge-cutoff hedging (20), filler (22), excessive hedging (23), generic conclusion (24).

Analyst B found 7 hits: AI vocabulary (7), copula avoidance (8), negative parallelism (9), rule of three (10), synonym cycling (11), false ranges (12).

Analyst C found 4 hits: em dashes (13), boldface (14), inline-header lists (15).

Phase 2: Consolidated report

Total: 25 findings (15 high, 7 medium, 3 low)

Phase 3: Rewrite

AI coding assistants can make you faster at the boring parts. Not everything. Definitely not architecture.

They're great at boilerplate: config files, test scaffolding, repetitive refactors. They're also great at sounding right while being wrong. I've accepted suggestions that compiled, passed lint, and still missed the point because I stopped paying attention.

People I talk to tend to land in two camps. Some use it like autocomplete for chores and review every line. Others disable it after it keeps suggesting patterns they don't want. Both feel reasonable.

The productivity metrics are slippery. GitHub can say Copilot users "accept 30% of suggestions," but acceptance isn't correctness, and correctness isn't value. If you don't have tests, you're basically guessing.

Phase 4: Self-audit

• Rhythm is fairly even across paragraphs -- acceptable but could use one longer sentence
• No remaining AI vocabulary, no chatbot artifacts, no formulaic structures
• Verdict: clean enough to finalize

Changes summary

• Removed chatbot artifacts ("Great question!", "I hope this helps!", "Let me know if...")
• Removed significance inflation ("testament", "pivotal moment", "evolving landscape")
• Removed promotional language ("groundbreaking", "nestled", "seamless, intuitive, and powerful")
• Removed vague attributions ("Industry observers")
• Removed -ing phrases ("underscoring", "highlighting", "reflecting")
• Removed negative parallelism ("It's not just X; it's Y")
• Removed rule-of-three and synonym cycling ("catalyst/partner/foundation")
• Removed false ranges ("from X to Y, from A to B")
• Removed em dashes, boldface headers, inline-header lists
• Removed copula avoidance ("serves as", "stands as") in favor of "is"/"are"
• Removed formulaic challenges section
• Removed knowledge-cutoff hedging, filler phrases, excessive hedging, generic conclusion
• Added personal voice, varied rhythm, specific observations

---

Behavioral notes

Short texts (< 3 paragraphs): Skip parallel analysis. Do a single-pass scan and rewrite yourself -- the overhead of 3 subagents isn't worth it for a few sentences.

Scoped requests ("just fix the em dashes"): Run only the relevant analyst (e.g., Analyst C for style patterns). Skip the full pipeline.

No findings: Tell the user the text reads naturally. Suggest soul-level improvements only if the writing is clean but lifeless.

Long texts (10+ paragraphs): Run the full pipeline. The rewriter may work section-by-section to maintain coherence across a long piece.

User provides context: If the user specifies the intended audience, register, or voice, pass that context to the rewriter subagent as additional instructions.

---

Reference

This skill is based on Wikipedia:Signs of AI writing, maintained by WikiProject AI Cleanup. The patterns documented there come from observations of thousands of instances of AI-generated text on Wikipedia.

Full pattern catalogs with word lists and before/after examples:

• references/content-and-communication.md
• references/language-and-grammar.md
• references/style-and-formatting.md