Kodownik 示範如何把個人寫作風格整理成 Claude Code 技能

It's eight in the morning. The coffee is still brewing. In the team channel someone asks why the pipeline is red. You already know: the runner's disk is read-only, the code is fine. You ask your agent to reply.

You get an essay. It opens with "I took a look." It has six sentences, four identifiers and the mechanism laid out step by step. It ends with "Let me know if you have any questions." Every word of it is true. None of it is you.

You add three words: in my words, concise, minimal. The second version is better. You send it. The next morning you do the same. And the next.

Psychology knows a man who traced a star seen in a mirror day after day, each time sure he was doing it for the first time. After brain surgery Henry Molaison could no longer form new memories of facts and events, yet his hand still learned the task. Your agent also meets you for the first time every morning. Only this time, you're the one repeating the exercise.

Those three words aren't a request. They're a memory that dies every morning.

The hand knows, the head doesn't

Try describing your style. You'll get something like "short, to the point, human." Everyone thinks that about themselves, and everyone reads those words differently. Your "short" is two sentences. Someone else's is a paragraph without an intro.

It's not your fault. Writing style lives in the same drawer as riding a bike. Your hand knows how you write; your head can't recite it.

But there's a side door. Recalling something from a blank page is hard. Recognising it when it's lying in front of you is much easier. The whole method rests on that: you don't ask yourself how you write. You lay cards on the table and look.

Twelve cards

Here's how we did it at Kodownik. First, material: two years of messages from the team chat and a few dozen comments from the issue tracker. Bots and private matters were dropped right away. Then twelve messages went on the table, shuffled, each with one question: mine or not mine?

You'd expect a simple sort. Something else happened. Among the "mine" were messages written by someone else. Among the "not mine" landed four of my own.

The question was "mine or not mine", but the answer turned out to be an answer to a different one: would I sign this?

The accepted ones shared a spine: the cause in one sentence, what's been done, and the next step with a named owner. The rejected ones fell into two families. Some were overloaded. Others said too little, and the shortest had two words and a question mark: "Anyone on this?"

Your style isn't how you write. It's how you write on your best day.

Two versions side by side

The details are settled with the same trick. For every doubt, two versions of the same message go on the table. The question isn't "which do you prefer", it's "which one do you send".

Options as prose or as a list? A list, and when there's one clear path, a note saying "this is the one I'd pick". To someone who decides rather than codes? The effect instead of the mechanism, and one request for a decision at the end. To a tester? What to check and what the fix doesn't change. Nine such pairs, nine decisions. None required describing myself.

The hand instead of the head

For the model, a style file is context, not a constraint. In Claude Code a skill is picked up when the model recognises a matching situation from the skill's description, and sometimes it won't. So next to it stands a gate: a PreToolUse hook that won't let a message out to the chat or the issue tracker until the agent has read the style. The gate doesn't ask the model's opinion. (As of the November 2026 docs, a PreToolUse hook can deny any tool call, MCP tools included.)

Before it was saved, the file went through a test: three real situations from the previous days, rewritten in the new style. All three passed without a single correction. Had any needed one, each correction would have been a missing rule.

Where this story overreaches

A style file isn't you. The model imitates the shape of a message. The voice will neatly wrap a wrong diagnosis, too. The content and the responsibility are still yours.

Twelve cards is not much. It's the calibration of one person in one conversation. Some rules are guesses, and the file labels them as such.

The gate checks that the style was read, not that it was followed.

Your turn

1. Gather traces. Messages to people from recent years. Drop the bots and private matters.
2. Lay out twelve cards. Mix yours with other people's. Answer only: mine or not mine.
3. Choose from pairs. Turn every doubt into two versions of the same message.
4. Write the portrait. One model sentence, a skeleton, the voice, a "never" list, rules per audience, "not like this → like this" examples. Separately: where each rule came from and which ones are guesses.
5. Let it write a letter. Three real situations from the past week. Every correction you make is a missing rule.
6. Put up the gate. A hook before sending that refuses until the style has been read.
7. Leave a door. A table of decisions with dates, kept in git.

To start, you can paste this to your agent:

I want to capture my style of writing to people as a skill. Walk me through a calibration:
1. Take my messages from [where] from the last two years, without bots and private matters.
2. Show me twelve, mixed with other people's. For each, ask only: mine or not mine.
3. From the accepted and rejected ones, extract a common skeleton. Show every doubt as two versions of the same message and let me choose.
4. Write SKILL.md (model sentence, skeleton, voice, never, audiences, "not like this → like this" examples) and SOURCES.md (where the rules came from, which are guesses).
5. Rewrite three of my real messages from the last week in the new style. Turn each of my corrections into a rule.

Henry Molaison could never say that he had practised the star. You can't say how you write. You don't have to. It's enough to recognise yourself on a card once.

And from tomorrow on, you won't have to tell your agent who you are.

Originally written in Polish for the Kodownik reading room: https://kodown1k.github.io/czytelnia/moimi-slowami/ — with illustrations and the first part of the story, "The jay, the nutcracker and CLAUDE.md".

Sources: Claude Code docs on skills and hooks; Henry Molaison (Wikipedia).