ihorh/jstyle

An opinionated prose style for your AI agent, so more of what you write gets read.
Install ·
An Example ·
How It Writes
Disclaimer · License
> Readers stay with your text. The skill makes the first
> lines count, and keeps each sentence readable in one pass.
> Cold reads find the gaps. A fresh reader, with none
> of the agent's context, points out what the agent can't see in its own text.
Agents write well now, so what goes wrong is small. A text opens with
background instead of what the reader came for. A sentence has to be read
twice. A paragraph's point arrives late. Each slip is minor, but together
they make prose that is good enough and slightly off. You feel it before you
can say why.
So do your readers. They take in text a few words at a time, and the first
few decide whether they stay. The skill aims to make those words count.
Your reader stays with you instead of asking their agent for a summary, and
more of what you write gets read.
> [!NOTE]
> The skill is for writing text that a person will read, such as
> documentation and manuals, code comments and docstrings, emails and
> conversations, many kinds of blog posts, and perhaps marketing copy.
It isn't for text that answers to a rulebook before a reader, such as legal
documents, government papers, or machinery specifications. It's also
unlikely to suit creative writing.
jstyle is a Claude Code plugin with one skill, /jstyle:write, which your
agent loads automatically when it writes for someone else. The repository also
holds a benchmark and other supporting material.
## An Example
One task from the first blind benchmark: Opus 5.5 updates a CSV tool's
README, so users learn how it handles very long rows before they run into
them.
Without the skill:
> A quoted field can contain newlines, as RFC 4180 allows. The field stays in
> one row. Each row is limited to --max-row-bytes, measured in UTF-8 bytes
> and not counting the newline at the end. The limit stops one stray " from
> turning the rest of the file into a single field.
> When a row goes over the limit, splitlane:
> 1. Keeps the first --max-row-bytes bytes of the row and writes them to the
> chunk as a normal row. The rest of the row is dropped, so the written row
> can have missing columns or an unclosed quote.
With the skill:
> splitlane cuts a row that is longer than --max-row-bytes and still writes it.
> It does not skip the row. The chunk gets the first --max-row-bytes bytes of
> the row, and splitlane drops the rest of the row up to the next line break. The
> cut usually falls in the middle of a field. The row that gets written can then
> have too few columns or an unclosed quote. An importer may reject it, or accept
> the shortened row without noticing.
The maintainer picked the version with the skill without knowing which was
which, because the other "lost my attention after mentioning RFC 4180." The
version without the skill is the best of its three tries, and the one with
the skill is its first. The
report has the rest
of the story: nine more tasks, and the reason behind every pick.
## Install
In a Claude Code terminal session, version 2.1.275 or later, run this,
confirm adding the marketplace, and choose a scope: