Open source

claude-style-patch is a one-line append to CLAUDE.md that rewrites how Claude writes prose, and picked up 126 stars in three days

September 11, 2026 at 2:35 AM PT

GitHub social card for the andrewroxby/claude-style-patch repository

Image: Andrew Roxby

Why it mattersA team that reads a lot of Claude output can drop the block into one file and get shorter sentences, plain words, and fewer announced-then-made points, without training a model or writing a custom prompt from scratch.

Andrew Roxby published claude-style-patch on GitHub on 8 September 2026, and the repository is at 126 stars three days later. The project is one file, STYLE.md, that a developer appends to their global ~/.claude/CLAUDE.md or to a project's own CLAUDE.md, licensed CC0 with no attribution required.

What the file is

Roxby writes in the README that the file is roughly 1,600 words, and that he arrived at it by naming, one by one, the specific tics in Claude's default house voice that were bothering him, and adding each as a rule to CLAUDE.md. He then re-ran conversations until the prose changed, and iterated. The rules span Claude Code and claude.ai, and he cites results from Opus 4.8 onwards.

The habits the file targets, in Roxby's own summary: Claude announces a point before making it, hinges sentences on a colon where the left side labels what the right side does, opens paragraphs with verbless fragments, and stacks compressed phrases against each other until a sentence has to be decoded rather than read. Roxby says these habits get worse as a conversation gets longer and more abstract, which is exactly when clear prose matters most.

The rules, and why they are stated as bans

Every rule in STYLE.md names the habit, gives an example of it, and gives the rewrite, because Roxby writes that Claude follows a concrete rule far better than a stated preference. The colon rule bans a colon followed by a clause except to introduce a literal list of three or more items, and requires the writer to turn the rest into two sentences or a clause joined by "because", "so", "but" or "and". The "Say It, Don't Announce It" rule bans signalling phrases such as "here's the thing" and "the key insight is", and treats verbless fragments as sentences as announcing by other means. The stacked compression rule asks the writer to keep verbs as verbs, use at most one metaphor per sentence, and never set two compressed units side by side. The proportion rule says a response ends when the content ends: no summarising and no engagement-bait closing questions.

Compliance is partial, and Roxby says so

Roxby is direct about the limits in the README. Compliance is good but not total, and it degrades on long threads. The Fable model, in his experience, follows the spec more consistently than Opus does, again more so on longer threads. The opening paragraph of the file tells Claude to re-check the rules when the material turns dense, which he says helps and does not fully solve the problem, and asking Claude directly to "hew closely to the response style instructions" before a long generation seems to work well on top of that.

The file works pasted into a Claude Project's custom instructions or into any system prompt, according to the README, so its use is not limited to Claude Code. The CC0 licence means a team can copy the rules into an internal style guide with no attribution.

Source

Source: Andrew Roxby

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

More from AI News

OKF Agent Memory gives coding agents a searchable memory with no embedding API

OKF Agent Memory stores agent memory as plain Markdown files in your repository and searches them with BM25 in under 300 microseconds, so there is no vector database and no embedding API to pay for.

Source: Hacker NewsOpen source

Agent-memory hits 52.9% on LongMemEval-S, 17 points ahead of MemCore, with Claude Code and Codex sharing one store

Agent-memory is an open source long-term memory runtime whose store is plain Markdown, whose index is a rebuildable SQLite cache, and whose paired benchmark on LongMemEval-S puts it at 52.9% against MemCore's 35.8%.

Source: GitHubDev tools

MonoCode puts eight coding agents in one desktop window and does not resell tokens

MonoCode is an MIT-licensed desktop app that runs Claude Code, Codex, Cursor, Grok Build, OpenCode, Pi, omp and fx through your own logins, and reached 623 stars in seventeen days.

Source: GitHubDev tools