How to stop Claude from sounding like Claude (the Unslop skill)

You can spot it in three seconds. The em dash where a comma should be. The "it's not just X, it's Y" construction. The sentence that starts with "In today's world" and ends with a summary nobody asked for. Readers have gotten so good at smelling AI-generated text that entire Reddit threads exist just for spotting the tells.

Anthropic knows this too, which is why Claude now ships with something called Skills — small, reusable instruction packages you can build once and reuse forever. One of the more clever community-built ones is called Unslop, and it exists for exactly one reason: to scrub the "AI smell" out of anything Claude writes for you.

Here's what it does, why it works better than just asking Claude to "sound more human," and how to install it in your own account in about five minutes.

What is a Claude skill, exactly?

A Claude Skill is a folder of instructions Claude loads automatically when it decides they're relevant to what you're asking. Anthropic's own documentation describes it plainly: Skills are folders of instructions, scripts, and resources that Claude loads dynamically to improve performance on specialized tasks, teaching Claude how to complete specific tasks in a repeatable way.

At the most basic level, a skill is just one file. Every skill consists of a directory containing at minimum a SKILL.md file, which must start with a YAML frontmatter to hold name and description fields, plus instructions for Claude or reference files, executable scripts, or tools.

That's the entire trick. No API keys, no code, no plugin marketplace approval process — just a text file with a name, a description, and instructions. Anthropic has open-sourced a whole library of these in the official Skills repository, covering everything from document formatting to frontend design systems.

Why does AI writing sound so recognizable?

There's actual research behind this, not just vibes. A 2026 study out of the University of Maryland and Google DeepMind analyzed over 61,000 texts and found something surprising: when AI text is edited to remove clichés and surface artifacts, detection based on structural features barely drops, from 95.5% to 93.9%.

In other words, swapping out "delve" for "explore" doesn't fool anyone. Every humanizer promises to make AI text sound human, and most of them just swap vocabulary, but the wordlist decays — "delve" peaked in 2023 and collapsed in 2025. Even punctuation habits shift as models update — GPT-5.1 suppresses em dashes — which means any tool relying purely on a static blocklist goes stale fast.

The real tells are structural, not lexical. That's the whole premise behind Unslop.

What does the Unslop skill actually check for?

Unslop is built around three tiers of detection, pulled partly from Wikipedia's well-known "Signs of AI writing" reference page and partly from patterns its creator noticed by hand. According to the developer's own breakdown: the detection patterns are organized into three tiers — dead giveaways like vocabulary clusters ("delve," "tapestry," "pivotal"), copulative avoidance ("serves as" instead of "is"), and formulaic significance framing ("is a testament to"); strong signals like promotional tone, dangling participle filler, synonym cycling, and vague attribution ("experts argue"); and stylistic tells like em dash overuse, hedge stacking, transition word spam, and sycophantic openers.

Altogether, it's built around a specific rule set. One community writeup puts it at a rigorous framework of 31 specific rules to eliminate significance inflation, formulaic structures, and robotic vocabulary while actively encouraging personality, varied rhythm, and concrete details. Instead of just flagging words, it pushes the opposite direction too — asking for opinions, rhythm variation, and first-person voice where it fits, which is precisely the stuff most AI writing avoids by default.

One subtlety worth knowing: the skill doesn't fire on every single occurrence of a flagged phrase. The hardest part of the scanner was calibrating false positives, so it ended up using density-based thresholds rather than per-occurrence flags for most patterns. That's why it can catch "this isn't just a tool, it's a mindset shift" without also nuking every legitimate use of an em dash in your draft.

How do you install Unslop in Claude?

You don't need Claude Code or a terminal for this — the fastest path works entirely inside claude.ai. Here's the process, based on Anthropic's own guidance for custom skills:

Step 1: Turn on code execution. Custom skills need this enabled first. Go to Settings > Capabilities and make sure Code Execution and File Creation are turned on — skills won't work without this.

Step 2: Open the Skills panel. Head to claude.ai/settings/customize-skills or navigate manually: open Settings, then look under the Customize heading in the left sidebar — not under "Settings" — and click Skills.

Step 3: Choose "Write skill instructions." Claude gives you three ways to add a skill — build one with Claude, write instructions directly, or upload a file. For something like Unslop, the text option is fastest: copy the skill text from a registry page and paste it into "Write skill instructions" — that's the fastest path and it needs no terminal, no file download, and no Claude Code install.

Step 4: Paste the instructions and name it. Grab the Unslop instructions from a source like Ricardo Decal's writeup and GitHub repo or the listing on claudeskills.info, paste the full text into the instructions box, and give it a clear name like "Unslop."

Step 5: Nail the description field. This part actually matters more than people expect. The description is the highest-leverage field in the entire form, because Claude reads skill descriptions to decide which skill applies to a given request — a description that only says what the skill is gives Claude nothing to match against. Instead, phrase it the way Anthropic recommends in its own guide: focus on capabilities and scenarios, not generic keywords, and include action verbs and clear use cases — instead of "document processing skill," write something like "extract tables from PDFs and convert to CSV format for data analysis workflows." For Unslop, something like "Detects and rewrites AI-sounding text — use when asked to edit, humanize, or clean up a draft" gives Claude an actual trigger to match against.

Step 6: Test it. Paste a paragraph of obviously AI-flavored text into a new chat and just say "unslop this." If nothing happens, the description probably needs to be more specific — test triggering and execution separately, and if skills don't activate, broaden your description and add use cases.

Can you build a skill from scratch instead of copying one?

Yes, and Anthropic built a tool specifically for this. It's called skill-creator, and it ships inside Claude itself. Per Anthropic's own resource guide: the skill-creator skill is built into Claude.ai and available for Claude Code, can generate skills from descriptions, and reviews and provides recommendations — just ask "Help me build a skill using skill-creator."

This is genuinely the easiest way to turn any recurring workflow into a reusable skill. Describe what you want ("scan my writing for AI patterns and rewrite anything that sounds robotic"), and Claude drafts the SKILL.md file, complete with a name and a triggering description, for you to review and save. You can also use it in reverse — ask "Review this skill and suggest improvements" and skill-creator can assess your existing skills for you.

There's also a hands-off option available to some plans: recording a skill is available on Pro, Max, and Team plans in Cowork in Claude for Mac — instead of writing a skill by hand, you can record yourself doing a task and let Claude build the skill from what it observes.

What are the limits worth knowing?

A few things to keep in mind before you build a stack of ten skills and expect Claude to juggle all of them perfectly.

First, skills are personal by default. Custom Skills are individual to each user — they are not shared organization-wide and cannot be centrally managed by admins on most plans, so if you're on a team, everyone needs to install Unslop separately for now.

Second, more isn't always better. Multiple skills can activate if they address different aspects of your request, so aim for single-purpose skills rather than one mega-skill trying to do everything. Keep Unslop focused on tone cleanup and build separate skills for formatting, research, or brand voice.

Third, remember Skills sit outside Anthropic's zero-data-retention guarantees. Agent Skills is not covered by ZDR arrangements — skill definitions and execution data are retained according to Anthropic's standard data retention policy. Fine for a writing-style skill, worth knowing if you're pasting anything sensitive into instructions.

Where do you go from here?

The bigger idea behind Unslop matters more than the specific skill. AI detection isn't really about banning "delve" anymore — it's about structure, rhythm, and the willingness to have an actual opinion in a sentence instead of hedging every claim into oatmeal. A skill like this doesn't make Claude a better writer on its own. It makes Claude apply your editorial standards consistently, every single time, without you having to type the same instructions into every new chat.

Once Unslop is installed, the next move is obvious: build a second skill that captures your actual voice — the phrases you overuse, the structure you like your intros to follow, the jokes you'd actually make. Pair that with Unslop's cleanup pass, and you've got something closer to a personal editor living inside Claude than a chatbot pretending to be one.