Skills

varied-voice: A writing style skill for reducing Claude's catchphrases

If you want to eliminate the recurring catchphrases and formulaic patterns in Claude's replies and turn the output into plain, information-dense prose with varied sentence lengths, varied-voice is the best fit; the caveat is that it is only a style-rewriting skill — its file, command, and network behavior is not stated officially, so for a first attempt use only pasted text and install it on officially supported platforms.

Official source · classicgoldbug/varied-voice

Install

cp claude-code/varied-voice.md ~/.claude/output-styles/

Who it's for

  • For anyone who reads Claude output regularly and is bothered by its recurring catchphrases and formulaic patterns: use it on any text that will be read by humans, replacing "Claudisms" with informative, ordinary prose — varied sentence lengths, plain connectives, and necessary hedges kept in their conventional places.
  • Officially, it is described as applicable to "anything a human will read," including articles you'll publish, documentation, and everyday replies.
  • When it's not a fit: if you care about fact-checking, formatting standards, or code quality, this is a writing-style skill and does not address those. For non-English text, or for details such as whether it triggers automatically on a specific platform, the official evidence says nothing — verify on your own before use. The entire skill is a single varied-voice/SKILL.md file (ten sections in total), follows the open Agent Skills standard, is pure Markdown, and is MIT-licensed.

First try in 3 minutes

The recommended main path installs it as a Claude Code output style; the commands can be copied directly:

  1. Prerequisite: Claude Code is installed locally, and this repository has been downloaded.
  2. Install location: copy claude-code/varied-voice.md from the repository into the ~/.claude/output-styles/ directory.
  3. Activate: run /output-style in Claude Code and select Varied voice. Official note: the change takes effect after /clear or in a new session.
  4. Success signal: give the agent a short piece of English prose — your own or the kind of formulaic text Claude produces (make clear it is text to be rewritten) — and ask to "rewrite it with varied-voice." What you can observe is output that becomes prose with varying sentence lengths and ordinary connective style, with formulaic phrases replaced.

This is a pure text style-rewriting task. Whether it reads or writes files, executes commands, or accesses the network on your machine is not stated in the evidence and is unknown — for your first attempt, use text pasted into the conversation, and don't hand it files you don't want touched.

Secondary method (officially supported, but not the main path): in the Claude app, download varied-voice.skill from the releases page, or zip the varied-voice folder yourself (the folder must be the root of the zip); enable code execution under Settings > Capabilities, then go to Customize > Skills, click + to upload the zip, and turn it on. To make it apply to every reply, add this line in your preferences (Settings > Profile): "Always load and follow the varied-voice skill before writing a reply." Pasting SKILL.md into another tool's custom instructions also works, but you lose automatic triggering.

Use cases

Case one: cleaning up a short piece of English prose. Paste a passage with obvious Claude-style formulas to the agent, state that it is text to be rewritten, and ask for a rewrite per varied-voice. The skill will apply its ten sections of rules to replace repeated phrases with ordinary prose. Within scope of expected results: varied sentence lengths, formulas gone; beyond that, no specific scoring or checking procedures are listed in the evidence and cannot be confirmed.

Case two: restyling every reply in Claude Code. After installing the output style via the main path and selecting it with /output-style, start a new session and simply ask questions normally — officially, the style changes the output writing style. What can be confirmed up to this point: the output style becomes Varied voice. Whether it also affects all other output types, such as tool-call reports or code comments, is not stated in the evidence.

Boundaries of use

  • File and command access: in the Claude Code path it works as an output style; whether it reads or writes files or executes commands is not stated officially and is unknown — don't infer zero risk from "it's just style."
  • Network access: no official statement; unknown.
  • Side effects and confirmation points: what is rewritten is the model's reply text; it does not directly modify files on disk because of that, unless the evidence says otherwise. The Claude app path requires enabling code execution under Settings > Capabilities — a switch that needs your explicit confirmation.
  • Runtime environment and client prerequisites: Claude Code (~/.claude/output-styles/ + /output-style, effective after /clear or a new session) and the Claude app (upload the zip under Customize > Skills). Note that skills load on demand; in the Claude app, the preference-setting instruction is needed for it to apply to every reply. Other platforms only support manually pasting SKILL.md into custom instructions, losing automatic triggering.
  • Official limitations: automatic triggering is only available in environments that support Skills; paste-based use does not trigger automatically.
  • Not yet confirmed: supported language range, whether it works with non-Claude agents, and specifics of the rewriting procedure — none of these are covered by the evidence; verify on a small scale before use.