Turn Your Writing Into a Claude Skill That Sounds Like You

Your voice is a dataset, and it is smaller than you think. Fifty thousand words you actually wrote, distilled into two hundred lines of rules and examples, is enough to make Claude produce drafts your readers can't reliably tell from yours. The delivery mechanism is a Claude Code skill: one SKILL.md file, loaded on demand, that rewires how the model writes for the duration of the task. This is the manual version of what installs.me automates. Here is the full pipeline.

Step 1: Build a corpus that is actually you

The single biggest quality lever is what goes in. Collect 30,000 to 80,000 words of prose where you controlled every sentence:

• Published blog posts and newsletters. Best signal per word.
• Long emails and memos you sent, unedited by anyone else.
• Talk transcripts and podcast appearances, lightly cleaned. These capture how you argue out loud, which is different from how you type.
• Investor updates, postmortems, strategy docs. These carry your decision-making style, not just your prose style.

Leave out, ruthlessly:

• Anything edited by committee. Marketing copy, press releases, legal-reviewed anything. It averages you into nobody.
• Ghostwritten or AI-assisted drafts. You would be cloning a clone.
• Slack noise. One-line reactions teach the model to be terse and useless.
• Your oldest writing, if your voice has moved. Weight the last three years.

Dump it into a folder as plain text or Markdown, one file per piece, filenames with dates. You will reuse this folder later as the skill's references/ directory.

Step 2: Distill voice rules, not vibes

Do not paste the corpus into a prompt and hope. Context windows reward compression. Run a distillation pass: give Claude the corpus in batches and ask for falsifiable rules, meaning rules a copyeditor could check with a highlighter.

Bad rule: "Direct and engaging tone." That describes everyone's aspirational LinkedIn bio.

Good rules look like:

• Opens with the conclusion, never with context-setting. First sentence is quotable in isolation.
• Median sentence under 15 words. Long sentences appear only to land a contrast.
• Uses second person when giving advice, first person when telling war stories.
• Never uses "leverage" as a verb, "journey", or "excited to announce".
• Disagrees by restating the other position fairly first, then one sharp sentence.
• Numbers over adjectives: "$30M exit", not "a significant exit".

Aim for 20 to 40 rules. Then extract anti-patterns separately: the words and moves you never make. Negative constraints do more work against AI slop than positive ones, because the model's default failure mode is regression to the generic mean.

Step 3: Name your frameworks

Voice gets you tone. Frameworks get you substance, and they are what make a persona skill useful for advice rather than just ghostwriting. Every experienced operator has 5 to 15 recurring mental models they reach for. Most are unnamed. Name them.

Go through the corpus and pull out every piece of repeated reasoning: how you evaluate a hire, when you kill a project, what you check before pricing anything. Write each as a named framework with a one-line trigger ("use when a founder asks about fundraising timing") and the actual steps. A named framework does two things a paragraph of advice can't: the model retrieves it reliably, and the output sounds like a person with positions instead of a search engine with manners.

Step 4: Write the SKILL.md

A skill is a directory with a SKILL.md at its root. The YAML frontmatter carries name and description, and the description is load-bearing: it is the trigger Claude matches against to decide when to load the skill.

---
name: yourname
description: Invoke when the user wants Claude to write, ghostwrite,
  or advise as <You>. Triggers on "in my voice", "as <name>",
  "draft this for my newsletter", or any first-person writing
  tied to <your blog / company / portfolio>.
---

# Writing and advising as <You>

## Who this is
Three sentences of biography with the concrete facts the model
should cite naturally.

## Voice rules
The 20-40 rules from Step 2.

## Never
The anti-pattern list.

## Frameworks
Each named framework, with its trigger condition.

## References
- references/corpus/ contains full original writing. Read 2-3
  relevant pieces before drafting anything longer than a tweet.

Move the raw corpus into references/corpus/. Skills load lazily, so the corpus costs nothing until the model actually needs to imitate a specific format, at which point it reads real examples instead of hallucinating your style from the rules alone. Rules set the boundaries; examples set the texture. You need both.

Step 5: Test the voice like it's code

The failure mode of every persona skill is drift back to default-Claude within three paragraphs. Test for it deliberately:

• Blind test. Generate three paragraphs on a topic you have written about, shuffle them with two real paragraphs of yours, and ask someone who reads your stuff to pick the fakes. Below 60% detection, you are done.
• Regression prompts. Keep 10 fixed prompts (a cold email, a hot take, advice to a founder, a eulogy for a dead project). Rerun them after every skill edit and diff the outputs. Skills regress exactly like code.
• Slop scan. Grep outputs for your banned list plus the universal tells: "delve", "it's worth noting", "in today's landscape", rhetorical questions as transitions. Every hit becomes a new line in the Never section.
• Disagreement test. Ask it something you have a contrarian position on. Generic-agreeable output means your frameworks are too thin, not your voice rules.

Shipping it

To make the skill installable by other people, wrap it in a plugin: a .claude-plugin/ directory with a marketplace.json, served from a git-backed source (plain relative HTTP paths in a marketplace will silently fail; the source must resolve over git). Then anyone runs /plugin marketplace add <url> followed by /plugin install <name>@<marketplace> and your voice is live in their terminal. Verified on Claude Code CLI 2.1.x.

The manual pipeline above takes a focused weekend. The corpus assembly is the tedious part, which is exactly the part worth automating.

Install a person

installs.me turns your files, calendar and calls into a Claude Code plugin that thinks like you. Anyone installs it with two commands:

/plugin marketplace add https://installs.me/lautaro
/plugin install lautaro@lautaro-installs