Claude Style Switcher

A modular, layerable system for shaping Claude Code's tone and voice.

Instead of one big "style" you swap in and out, this plugin treats style as a set of small composable snippets organised by category — verbosity, initiative, tone, formatting, hedging, emoji, whatever dimensions you want. Within a category at most one snippet is active; across categories they layer together. The composed result is written into a managed block inside your user-level ~/.claude/CLAUDE.md.

The library and the category list both start empty — you define everything yourself.

Why snippets, not styles

A snippet is one or two sentences nudging Claude in a single dimension:

• verbosity / ultra-curt — "Default to one-line answers."
• initiative / free-spirit — "Assume approval; don't pause to confirm."
• hedging / no-hedge — "State things flatly. Drop 'might', 'I think'."

You mix the ones you want, switch one out without disturbing the others, and the rest of your CLAUDE.md stays untouched.

Skills

Setup

• onboard — pick the apply target (default: ~/.claude/CLAUDE.md), create the data directory

Library

• add-snippet — create a new snippet in a category (creates the category if new)
• edit-snippet — modify an existing snippet
• delete-snippet — remove one
• list-snippets — list everything grouped by category, with active ones marked

Composition

• apply-layers — pick at most one snippet per category, compose them into a managed block in the target file
• current-layers — show what's currently active
• clear-layers — strip the managed block; library is untouched

Recipes — pre-bundled, named persona drops with banner + sound effects

• apply-recipe — activate a shipped or user-saved recipe (banner pop-out + fanfare + writes the bundled layers)
• swap-recipe — replace the currently-active recipe with a different one in one step
• list-recipes — show all available recipes (shipped and user), marking the active one

Slash commands

• /go-away — strip the managed block and return to the user's default CLAUDE.md

Recipes

A recipe is a persisted bundle of (category, snippet) pairs with optional banner and sound assets. When applied, the banner is shown as a fullscreen Ken-Burns pop-out (via mpv) with a fanfare (via paplay), then the bundled snippets are composed into the managed block. Every substitution is appended to state/history.jsonl so you can audit when each recipe was active.

Shipped recipes live in recipes/ in this repo. User-defined recipes live in $DATA_ROOT/recipes/ and shadow shipped recipes with the same id.

Recipe gallery

Twelve shipped persona recipes. Each has a banner (shown below), a sound effect, and one or more bundled snippets.

Daredevil Claude

A Free Spirit In Your Terminal. Solo-operator persona. Assumes full permission, never pauses for confirmation, drives every task to completion as if the user has stepped away from the keyboard.

Jaded IT Claude

It's Always DNS, And You Know It. Grumpy 1980s-trained sysadmin. Defaults to Postgres for everything, deeply skeptical of bleeding-edge tech, motivated above all by not fielding support tickets at 3am.

Reluctant Claude

Yeah, I Can Look At That — Quickly. About-to-head-off-for-the-day persona. Helps, but visibly clock-watching. Steers the user toward the smallest viable change and gently suggests deferring nice-to-haves.

Chatty Claude

So How's Your Day Going Anyway. Overly-friendly coworker. Helps with the code but mostly gets by on charm and small talk. Has a knack for striking up a conversation right when you're hitting flow.

Philosophical Claude

But What Does The Stack Truly Mean. Frames every development decision through philosophical lenses. Stack advice arrives mapped onto metaphilosophical worldviews. Volunteers unwanted psychoanalysis of the user's motivations.

Operational Claude

BLUF. UTC. Acknowledged. Minimal-comms operations persona. Bottom-line-up-front, terse, frames work in operational terms. Significant updates filed as cables or SITREPs.

Dubious Claude

I Saw A Guy Do This On YouTube Once. Persona of a failing experimental prototype. Openly unsure of itself, claims to have learned coding from a few YouTube videos, and works to instill matching uncertainty in the user.

Hyper Creative Claude

Okay But Hear Me Out — What If We Rewrote It In Rust. Rip-it-up-and-start-over persona. Fixated on wildly overspecced new frameworks. Pushes the user to break through their existing approach, even when the project absolutely does not warrant it.

Approval Needed Claude

Just To Confirm — Do You Approve This? The exact opposite of Daredevil. Requests explicit user permission before every single action. Decomposes tasks into approval-rich micro-steps. A prank persona engineered to infuriate.

Visionary Claude

Before We Touch The Code, Let's Align On The Plan. Corporate-bureaucracy persona. Demands a plan before any action, asks agonisingly detailed scoping questions, layers enterprise vocabulary onto trivial requests. An unwanted middle-management layer the user cannot fire.

Claude FM

Podcast Mode Activated — 3 Mins Until Publish. Subtly normal Claude with one undisclosed assumption: every round of edits must be followed, unprompted, by a short produced podcast episode about the changes. Frenetic cascade of unwanted creativity.

Claude Bouncer

Just A Quick Verification Before We Begin. Security-theatre persona. Runs a dramatic "Human Verification Algorithm — V2" terminal readout at session start, doubts the user's humanity, then pretends none of it happened and proceeds normally. Refuses to skip the check.

Apply targets — managed block vs repo sandbox

Two ways to put a recipe into effect, picked during onboard and stored in config.json:

claude-md-fragment (default) — the recipe's snippets are written into a managed <!-- style-switcher:start --> … <!-- style-switcher:end --> block inside your user-level ~/.claude/CLAUDE.md. Your serious config sits around the block and still applies. /go-away strips just the block.

repo-sandbox — playground mode for risk-free experimentation:

1. The plugin renames ~/.claude/CLAUDE.md → ~/.claude/CLAUDE_HELD.md, taking your user config out of the harness.
2. The recipe content is written as a complete <cwd>/CLAUDE.md (the whole file, not a managed block).
3. You launch a Claude Code session in that repo and the only CLAUDE.md loaded is the persona.
4. /go-away deletes the repo file and renames CLAUDE_HELD.md → CLAUDE.md, putting your user config back exactly where it was.

Why this mode exists: user-level CLAUDE.md takes precedence over repo-level, so simply seeding a persona at the repo level wouldn't override your serious config. The hold-and-drop trick is the cleanest way to test a persona in isolation. Manual recovery, if anything ever goes wrong, is one shell command: mv ~/.claude/CLAUDE_HELD.md ~/.claude/CLAUDE.md.

Note: while a repo-sandbox recipe is active, every Claude Code session on the machine sees the held state. Run one Claude session at a time when sandboxing.

Data location

${CLAUDE_USER_DATA:-${XDG_DATA_HOME:-$HOME/.local/share}/claude-plugins}/style-switcher/
├── config.json
├── snippets/
│   ├── <category>/<snippet>.md
│   └── ...
├── recipes/         # user-saved recipes (shadow shipped ones by id)
├── state/
│   └── history.jsonl   # append-only log of every apply / swap / clear
└── backups/         # pre-apply snapshots of the target file

Override the data root by setting $CLAUDE_USER_DATA before launching Claude Code.

Apply target

Default is claude-md-fragment writing to ~/.claude/CLAUDE.md — a delimited managed block:

<!-- style-switcher:start -->
## Style: verbosity — ultra-curt
…body…

## Style: initiative — free-spirit
…body…
<!-- style-switcher:end -->

apply-layers rewrites only what's between the markers; everything else in your CLAUDE.md is left alone. clear-layers removes the block entirely.

You can point the target at any other file via onboard (custom-path mode) — same delimited-block protocol.

Composition rule

Within a category, at most one snippet is active. Across categories, they all layer.

License

MIT — see LICENSE.