skill-creator

my-style-gen

Your agent writes well. It just doesn't write like you.

So the code it writes, the deck it drafts, the spec it fills in... all in a stranger's idiom: slower to read, deaf to the shorthand you actually use, and not quite yours by the time it ships. This fixes all three, once.

Point it at work you already produced: git repositories, specs, decks, spreadsheets, contracts, notes. It reads each source, works out how you work and how you write, and installs a skill called my-style that every future agent session follows.

• Faster review — the work and the prose explaining it read the way you write, and need fewer corrections
• Shared language — your vocabulary and shorthand travel with the skill, so a terse request is decoded the way a long-time colleague would decode it
• Personal touch — what you make with an agent still reads like your work

Works for any line of work, not just engineering (an engineer's sources are repositories, a product manager's are specs and decks, a lawyer's are filings — same method, same output). And it adopts the style, not the weakness: a modern model already writes a cleaner first draft than most of us, so each analysis agent filters as it reads. The goal is legibility, not forgery.

What it looks like

Prompt

Explain why this can charge a customer twice, and what we should change. (followed by a ten-line Go retry loop)

Response

Generated my-style skill

<!-- SKILL.md -->
- **State the mechanism, not the impression.** Name the machinery that makes the claim
  checkable. "Bounded by the queue depth once the consumers saturate", not "it scales
  better".
- **Put the reason in a parenthetical, not in a sentence of its own.** He compresses
  justification rather than narrating it. "Moved the publisher outside the transaction
  (so a failed publish can't roll back a committed order)" - not a paragraph explaining
  transaction semantics.

<!-- references/voice.md -->
- Problem before solution, always. Open by stating what the alternatives fail at, then
  the goal - never with your own solution first.

Install

One command, any agent — Claude Code, Copilot, Codex, Cursor, Windsurf and 75 others:

npx skills add marcotrinelli/my-style-gen -g -a claude-code

That installs it for every project. Drop -g to install into the current project only.

Then ask for it in plain words:

build a style skill from my work

Other ways to install

Claude Code, as a plugin:

/plugin marketplace add marcotrinelli/my-style-gen
/plugin install my-style-gen@marcotrinelli

By hand: clone and copy skills/my-style-gen into ~/.claude/skills/, ~/.agents/skills/, or wherever your agent keeps its skills.

Companion skills

my-style-gen works on its own, but hands off to and leans on a few skills that live upstream in anthropics/skills. Install the ones that match your sources:

All five in one go:

npx skills add anthropics/skills -g -a claude-code --skill skill-creator docx pptx pdf xlsx

In Claude Code skill-creator is also carried by the official marketplace, so /plugin install skill-creator@claude-plugins-official works just as well.

Without the format skills the analysis still runs — it falls back to unzip, pdftotext, pandoc or markitdown — but the read is lossier, and tracked changes and comments are where the best anti-pattern evidence lives. Plain text and git repositories need none of them.

Use

Asking in plain words is enough. Or give it what it needs up front — what you do, and where your material is:

/my-style-gen
/my-style-gen backend engineer ~/src/api ~/src/cli
/my-style-gen product manager ~/Documents/specs ~/Documents/decks
/my-style-gen patent attorney ~/matters/filings
/my-style-gen https://github.com/me/project

Whatever is missing, it asks for in one message — what you do, which sources to read, how to tell your material from everyone else's, your name and pronouns — then goes.

[!IMPORTANT] Point it at work you actually wrote. Material an agent produced end to end carries the model's idiom, not yours, so feeding it back gives you a skill describing the default you were trying to get away from.

Reading the sources is the expensive part (one agentic exploration per source), so prefer two or three substantial sources over many thin ones. Sources from unrelated contexts are worth more than two from the same team — agreement across them is what separates your habits from the house style.

The result installs as my-style, under ~/.claude/skills/ by default. Ask for another name where two people's skills have to sit side by side. If it lands inside a git repository, its path is added to that repository's .gitignore, because it describes you and not the project.

How it works

Three steps, one agent run per source. No tool to install, no state kept between runs.

Voice, vocabulary and standards carry the most weight. The first two serve legibility