mcp-overleaf-server: Skill for Claude Code

Humane Writing

Why this matters

A person reading a cover letter or a follow-up email isn't grading it against a rubric — they're forming an impression of whether they'd want to work with this person. AI-generated prose has a recognizable rhythm: it's evenly balanced, reaches for the same handful of intensifiers, and structures almost every paragraph the same way. None of that is wrong, exactly, but a reader who's seen a thousand cover letters this year notices the pattern before they notice the content. The goal here isn't to trick anyone — it's to write the way the user would actually write if they had the time and clarity to say it well themselves.

The tells

These are the patterns that make writing read as machine-generated. None of them are wrong in isolation — the problem is that a model reaches for all of them, all the time, and the reader feels the sameness.

Em-dash overload. This is the single most common tell, and the easiest one to fall back into out of habit — it's the default connector a model reaches for when it hasn't decided whether two clauses want a period, a comma, or a colon. Treat the default as zero, not "one or two." Before finalizing any document, go back through the actual draft and find every em dash specifically — don't rely on having "written carefully" the first time through, since the habit is exactly what needs catching. Replace each one with a period, a comma, or a colon depending on what the sentence needs. Only leave one in if the interruption is genuinely the point — a real aside or a hard pivot mid-thought, not just a place a pause felt natural. More than one in a short document is a sign the em dash is being used as filler rather than a deliberate choice.

The "not just X, it's Y" contrast frame. "This isn't just a résumé tool — it's a philosophy on how work should be tailored." This construction sounds insightful the first time and hollow the fifth. If the contrast isn't doing real work, cut it and just state the thing.