$ cat problem.md
The Problem
AI-generated writing has a distinct voice that screams, "Robots wrote me!" You know it when you see it. This includes:
- Formulaic openings and ritual conclusions
- Excessive hedging and meta-commentary
- Corporate buzzwords and empty superlatives
- Lack of authentic voice or personality
The problem isn't that AI can't write well on its own. The problem is that AI is defaulting to outputs deemed "helpful" rather than genuine or human. You end up with technically correct but soulless prose.
$ cat solution.md
The Solution
I wanted to test out Claude skills and eliminate AI writing tells. I use AI as a tool. If it's going to help me write, I need it to be as human-sounding as possible. So I built a Claude skill that detects and eliminates AI writing tells. The skill includes:
- Pattern detection: Identifies formulaic structures, puffery, and robotic phrasing
- Active revision: Rewrites flagged sections with authentic, direct language
- Voice preservation: Maintains the core message while removing AI tells. I often dictate so Claude understands my "voice". I want that preserved.
- Iterative refinement: Can be run multiple times until the output sounds human
The skill works by comparing text against patterns common in AI-generated writing, then applying principles of encyclopedic prose.
$ cat stack.md
Technical Stack
This is a Claude skill - a reusable prompt template that guides Claude's behavior:
- Claude Skills framework: Built as a portable .md file that can be invoked in any Claude interface
- Pattern library: Documented list of AI writing tells to detect and remove
- Rewriting rules: Guidelines for transforming flagged text into natural prose
- Anti-AI Writing Guide: Core principles adapted into executable instructions
The skill is version-controlled and can be shared across teams. When Claude processes text with this skill active, it automatically applies the anti-AI writing principles.
$ _
Example Transformation
❌ Before (AI tells):
"It's worth noting that implementing this feature could potentially provide significant value to users. Many industry experts believe that this approach may offer meaningful benefits, though results can vary depending on specific circumstances."
✅ After (human voice):
"This feature solves the user's core problem; reducing time spent on manual data entry. Early tests show 60% faster completion rates."
Notice the difference? The "after" version is direct, specific, and confident. No hedging, no empty phrases, no meta-commentary about what experts think.
Using the Skill
The skill can be invoked in any Claude interface:
Claude Code (CLI):
@anti-ai-writing [your text]
Claude Desktop:
Use skill: anti-ai-writing
Once active, the skill processes any text you provide and returns a rewritten version with AI tells removed. You can iterate multiple times until you're happy with the voice.
Want to build your own skill? Read the full tutorial