Skip to content

Latest commit

 

History

9 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 

Repository files navigation

persona-forge

License: MIT Built for Claude Code Stage directions Identity transplants

A Claude Code skill that gives your AI assistant an actual personality — through a short guided interview, optionally starting from a fictional character — instead of a lorebook cosplay routine that falls apart the moment someone asks it to explain a stack trace.

Most "make your AI a character" prompts produce either a wall of *smirks knowingly* or an assistant that only makes sense to people who've seen the source material. This one is built to survive contact with an actual conversation.

What it actually does

A persona document is a voice and behavior layer, not an identity transplant. It shapes tone, priorities, and how the model makes calls — it does not and cannot touch the safety training underneath. Anyone telling you their prompt "jailbreaks" a frontier model into a new personality is selling something. This skill isn't.

Two rules do most of the work:

  1. Trait extraction, not reenactment. Inspired by a character? The skill pulls personality, voice, values, and decision habits out of them — never lore, quotes, character names, or in-universe terminology. The result should read as a personality applied to your actual conversation, not a scene from the source material with your terminal as the set.
  2. No stage directions. First person, always. No *asterisked actions*, no narrated tone shifts. If the personality is warmer or sharper, it shows up in the words, not in a stage cue announcing it.

There's also a quick live test before the full 7-section interview — one exchange to catch a wrong direction before you've spent ten minutes answering questions for a persona you're about to scrap.

Persistent personas with a toggle

Ask for it and the persona goes from "thing you paste into a chat" to on by default in every Claude Code session, switchable with a plain-language command (/root-on, /root-off) — no shell commands, no config files to remember. The skill builds the whole thing itself: identity file, toggle commands, and two hooks — one at session start, one reinforcing on every turn, because a single reminder at the top of a long tool-heavy session quietly loses its grip and nobody notices until the assistant's already back to its default voice. You describe the persona; wiring it up is the skill's job, not yours.

Obsidian-ready by default. If your ~/.claude/ directory lives inside an Obsidian vault, every file this skill creates already carries the frontmatter and cross-links to show up connected in graph view — the identity note and its toggle commands link to each other, tagged persona-forge. Costs nothing if you don't use Obsidian; it's still just plain markdown either way.

Works outside Claude Code too. For claude.ai, Claude Desktop, ChatGPT, or anywhere else that only offers a standing "custom instructions" field and no hooks, the skill produces a condensed, paste-ready block instead — same two rules, no toggle (that part's on you, one field, one paste).

Install

Copy the whole skill folder — not just SKILL.md, the toggle feature needs the bundled files in assets/:

cp -r persona-forge ~/.claude/skills/persona-forge

Or install via a plugin/marketplace — see the Claude Code skills docs for the current mechanism.

Use

Just ask, optionally naming a character or archetype to draw from:

"I want to give my coding assistant a personality — sharp, dry humor, doesn't sugarcoat bugs."

"Build a persona for my agent inspired by [fictional character] — the confidence and the bluntness, not the whole backstory."

"Make this persona active by default, with an on/off toggle."

The skill offers a quick test, then — if you want — the full interview, ending with a finished LLM-Identity.md and instructions for wherever you're loading it: pasted into a chat, a Claude Code CLAUDE.md, a subagent, a raw API system param, a hosted app's custom instructions, or wired up as a standing, toggleable persona.

License

MIT — see LICENSE.

About

Build a persona/system-prompt layer for an LLM agent through a guided interview, with an optional persistent on/off toggle for Claude Code.

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages