Model-Specific Skills: Writing for Claude, GPT and Gemini

The AI Voice Problem: Why Your Drafts Sound Generic

Ask a model to write in your voice and it will happily oblige — in a voice that belongs to nobody. That's the paradox practitioners keep running into: AI is perfectly consistent from the start, and that consistency is exactly why its output sounds generic. As Every.to's style-guide research puts it, the real challenge is pushing a model toward the idiosyncrasies that make writing belong to a particular person.

This isn't just a hunch. A peer-reviewed longitudinal study of 4,820 student reports written between 2016 and 2025 found that GPT imposes a measurable stylistic fingerprint on text. Post-ChatGPT writing became more formal — higher lexical density, more nominalisation — and more positive in sentiment, a documented "positivity bias" that appeared regardless of topic or how weak the underlying results were.

Critically, those shifts produced no improvement in quality. The same study found the stylistic changes corresponded with no better grades or feedback, raising what the researchers call questions about AI's impact on voice, creativity, and critical thinking. Polish, it turns out, is not the same as substance.

Detection specialists and practitioners have catalogued the tells that give AI drafts away:

• High-frequency AI words: "delve," "tapestry," "leverage," "seamless," "meticulous."
• Structural uniformity — monotonous sentence lengths, uniform paragraphs, list-like organization.
• Heavy em dash use; one writer estimates roughly 90% of AI-inserted em dashes can simply be cut.
• Buzzword framing and formulaic moves like "It's not just X. It's also Y," plus cliché openers about "today's fast-paced digital landscape."

Some of these patterns are almost comically specific. According to Pangram's analysis of AI writing patterns, 60–70% of names in AI-generated articles from ChatGPT and Claude are either "Emily" or "Sarah."

Models also differ in how strongly they impose a voice. A vendor comparison perceives Claude as producing the most natural, least "AI-sounding" prose with precise style adherence, ChatGPT as tending toward formulaic structures, and Gemini as competent but lacking voice adaptability — though these are directional perceptions, not verified benchmarks.

The practical takeaway: don't fight the fingerprint by re-pasting instructions into every chat. The scalable fix is a concrete style guide — specific sentence patterns, examples, and above all anti-patterns — enforced through reusable systems like Claude Projects, custom GPTs, or agent skills. Agent skills (SKILL.md instruction packages) are exactly that Level-3 mechanism, and SkillzBench is building a free directory and free MCP connector so you can discover, verify, and test writing-style skills across Claude Code, Codex, Cursor, and Gemini CLI — with security vetting, which matters given published research finding ~26% of skills carry at least one vulnerability.

AI cannot replace your point of view. But left unmanaged, it will bury it.

Model-Specific Traits: What the Data Says

Every AI model leaves a fingerprint on your prose — and the evidence for how strong those fingerprints are varies dramatically between GPT, Claude, and Gemini. Knowing which claims rest on peer-reviewed data and which are vendor perception is half the battle when you're choosing a writing assistant.

GPT has the strongest empirical track record of stylistic influence. A peer-reviewed longitudinal study of 4,820 student reports found that ChatGPT-associated lexical markers like "delve" and "intricate" surged in student writing until 2024, then declined — possibly because students started masking AI traces. The same study documented that GPT makes writing more formal, more positive, and lexically denser regardless of topic.

The catch: those stylistic shifts produced no improvement in grades or feedback. GPT polishes the surface without changing the substance, and its documented positivity bias raises questions about impact on voice and critical thinking. A rewriting experiment confirmed causality — GPT-rewritten pre-ChatGPT texts matched post-ChatGPT style almost exactly.

Claude and Gemini are a different story. The evidence here is thinner, and honesty matters: model-comparison claims rest largely on a single vendor comparison, not verified benchmarks. That source positions Claude as producing "the most natural, least 'AI-sounding' prose" with precise style adherence, ChatGPT as tending toward formulaic structures, and Gemini as competent but weaker at creative tone. Gemini's practical advantage is its context window — up to 1M tokens versus Claude's 200K and ChatGPT's 128K, per the same vendor data.

So what separates verified data from perception right now?

• Verified: GPT's formality, positivity bias, and lexical markers — peer-reviewed, pre-registered, 4,820 documents
• Directional: Claude's prose naturalness and Gemini's long-context strength — vendor-sourced perception, unverified versions and figures
• Corroborated across sources: AI tells like em dashes, buzzwords, and uniform paragraph structure appear in detection research and practitioner accounts alike

One quirky data point worth knowing: detection research found that 60–70% of names in AI-generated articles from ChatGPT and Claude are either "Emily" or "Sarah." Small tells add up.

The practical implication is that writing skills — reusable instruction packages that enforce your voice and blacklist these tells — work differently per model. SkillzBench is building exactly this kind of evidence base for agent skills: sandboxed benchmarking that measures whether a skill actually lifts output, so you're not relying on vendor claims. Until verified head-to-head writing benchmarks exist, treat Claude and Gemini strengths as informed perception, and GPT's fingerprints as fact.

Building Your AI Style Guide: The Anti-Pattern Approach

The most useful AI style guide isn't a list of adjectives you like — it's a list of behaviors you refuse to tolerate. Vague descriptors like "clear" or "conversational" give a model nothing to work with, because a model is perfectly consistent from the start; that's exactly why it sounds like nobody. The fix is an operating manual built on concrete examples and, above all, anti-patterns.

Practitioners who build these guides report that the anti-pattern section is often the highest-value part, and that examples paired with rules almost always beat rules alone. A practical blacklist looks like this:

• Ban the tell-tale vocabulary: "delve," "tapestry," "leverage," "unlock," "empower," "elevate."
• Cut the em dash habit — one practitioner estimates roughly 90% of AI-inserted em dashes can be replaced by commas.
• Kill formulaic frames like "It's not just X. It's also Y" and cliché openers such as "In today's fast-paced digital landscape."
• Force varied sentence and paragraph lengths, since uniform structure is one of the clearest AI tells.
• Preserve a point of view — AI can't replace a writer's perspective, and that remains the most reliable human marker.

The stakes are measurable, not cosmetic. A peer-reviewed study of 4,820 student reports found that GPT-driven writing became more formal and more positive in sentiment, with lexical markers like "delve" surging until 2024 — yet those stylistic shifts produced no improvement in grades or feedback. That is the trap: polish isn't quality. GPT's documented positivity bias raises real questions about voice, creativity, and critical thinking, and voluntary disclosure failed entirely in that dataset.

Model choice compounds the problem. A vendor comparison describes Claude as producing the most natural, style-adherent prose, ChatGPT as leaning formulaic, and Gemini as competent but less voice-adaptable — directional perceptions, not verified benchmarks. So write for the model in front of you, and counter its defaults explicitly.

The scalable move is to stop pasting guides into chat windows and treat them as reusable instruction packages instead. That's where agent skills fit naturally: SKILL.md-style packages for Claude Code, Codex, Cursor, and Gemini CLI agents can carry your anti-patterns automatically. SkillzBench is building a free directory and MCP connector to discover, verify, and benchmark those skills — worth watching given published research finding roughly 26% of skills carry at least one vulnerability.

From Chat Paste to Reusable Skills

Pasting a style guide into a chat window works — once. Do it a hundred times across a hundred conversations, and it becomes the kind of manual drudgery AI was supposed to eliminate. The practitioners at Every.to describe this as a three-level maturity model: paste-in-chat guides at Level 1, Claude Projects or custom GPTs that load your guide automatically at Level 2, and skill or plugin systems with automated checks at Level 3.

That third level is where writing style enforcement becomes genuinely scalable. Instead of re-explaining your anti-patterns every session, you package them once as an agent skill — a structured instruction file the model reads before it writes. The same anti-pattern blacklist that kills the "delve" and "tapestry" vocabulary, the em dash overuse, and the "It's not just X. It's also Y" framing gets applied consistently, whether you're drafting in Claude Code, Codex, Cursor, or Gemini CLI.

There's a practical reason to systematize this rather than trust raw model output. The peer-reviewed longitudinal study of 4,820 student reports found that GPT's stylistic fingerprints — the formality, the positivity bias, the lexical markers — appeared even when writers tried to mask them, and none of the polish improved actual quality. Style drift isn't a one-time fix; it's a recurring risk that needs standing guardrails.

But adopting third-party skills carries its own risk. Published research has found that roughly 26% of agent skills carry at least one vulnerability, so "found on GitHub" is not a security standard. Before loading any writing-style skill into your workflow, you want to know where it came from, who maintains it, and whether it actually helps.

That's the gap SkillzBench is being built to close. Its free directory will let you discover writing and style skills across agents, with every version pinned to a commit SHA, license-checked, security-scanned, and benchmarked in a sandbox — testing each skill with and without it on the same model, then publishing the Lift in percentage points so you can see whether a "Claude tone" skill genuinely moves the needle. A free MCP connector will let you toggle skills and tools per agent, so each agent only sees what you've chosen for it.

A few things to check when you evaluate any writing skill:

• Does it include concrete anti-patterns and examples, not vague instructions like "be conversational"?
• Is the version pinned and traceable to a specific commit?
• Has it been security-scanned, given that ~26% of skills carry vulnerabilities?
• Is there measured evidence it improves output, not just a five-star rating?

Reusable skills beat repeated pasting — but only if they're verified before they touch your workflow. SkillzBench is pre-launch, with its directory, benchmarks, and MCP connector in active development; you can join the waitlist now and be first to test writing skills across Claude, GPT, and Gemini when it ships.

ctaText: Join the SkillzBench waitlist — free skills, verified and benchmarked before launch.

socialProofText: SkillzBench is free for everything: the directory, MCP connector, benchmarks, and alerts.