The fingerprint
Every model writes the same way. The em dash pairs, the “not X, but Y” pivot, the sentence that ends by underscoring its own importance. Wikipedia’s AI Cleanup project has catalogued these tells across thousands of flagged articles, and once you’ve read the list you can’t unsee it. I was hand-editing the same patterns out of everything Claude drafted for me, including the copy on this site, so I turned the editing pass into a plugin.
This isn't just a design system — it's a journey. By leveraging a robust set of shared tokens, we unlocked seamless collaboration across the entire ecosystem, highlighting the transformative power of a common language.
The design system started as a spreadsheet of hex codes. Once the tokens lived in code, designers stopped pasting values into Slack and the front end stopped drifting. The shared language did the rest.
Guidelines outrank the model
mimesis (μίμησις, imitation) is six skills with one thesis: the quality lives in a versioned corpus of tells and a deterministic linter, not in whichever model happens to be writing. The linter is a single Python script with no dependencies. It flags em dashes as blockers, catches the en-dash and double-hyphen workarounds, and scores prose for flat rhythm and narrow vocabulary. The kill list is versioned by model era because the tells drift. There’s also a compile skill for the opposite job, writing dense instructions for a machine reader, where the tells are fine.
Jane is a passionate, results-driven engineer with a proven track record of delivering innovative, scalable solutions that empower teams to achieve their full potential.
Jane builds payment infrastructure. She rebuilt the reconciliation system twice, once wrong, once less wrong, and writes about what the second attempt taught her. Based in Edinburgh.
Imitating a person
There’s an obvious irony in asking Claude to remove Claude’s own tells, which is exactly why the rules live in files the model can’t argue with. The linter false-positives sometimes, and I’ve made peace with that; it reports, a human decides. What I haven’t made peace with is how many em dashes I used to write myself. Apparently the machine learned them from us.