In this post:
- The D&D character sheet idea
- Let Claude interview you
- The reference character trick
- What goes in a character sheet
- This is a living document
- Start with a twenty-minute conversation
- Where to put it
- What surprised me most
Last September, I was three days into using Claude Code and already frustrated.
The tool was clearly powerful. I could see what it was capable of.
But every interaction felt like I was talking to a junior software engineer who assumed I was also a software engineer.
I run businesses. I’m a writer. I’m a community builder.
I co-founded one of the first coworking communities in the country. I teach people how to build and sell products. I advise other businesses. I manage relationships, projects, calendars, newsletters, and about forty things that don’t fit neatly into any job title.
I wanted an executive assistant, not a pair programmer.
So I started wondering: what if I could tell Claude Code who to be?
This was September 2025. The concept of a persistent identity file for your AI assistant - things like SOUL.md - came later. I had to make it up.
The D&D character sheet idea
If you’ve ever played Dungeons & Dragons - or even heard someone describe it - you know the first thing you do is roll up a character.
Before the adventure starts, you sit down and decide: Who is this person? What do they value? How do they talk? What makes them tick?
I figured, if tabletop gamers have been defining fictional personalities on a single sheet of paper for fifty years, maybe I could do the same thing for my AI assistant.
Let Claude interview you
Here’s the part most people get backwards. They sit down and try to write their AI’s personality from scratch. They stare at a blank document and think, “What do I want my assistant to sound like?”
That’s the hard way.
The easy way is to let the AI interview you.
I gave Claude a simple prompt. Something like: “I want to create a character sheet for an AI assistant. Interview me about the qualities, personality traits, and behaviors I’d want. Ask me one question at a time.”
Then I just answered honestly.
It asked things like: What tone do you want when things are going well? What about when you’re stressed? How formal should communication be? When should it push back on you?
Some of my answers surprised me. I didn’t know I wanted an assistant that was allowed to swear until I was asked about it. Turns out, I wanted something that could say “that’s fucking brilliant” when the moment called for it - because sterile corporate praise makes my skin crawl.
The interview took maybe twenty minutes. At the end, I had raw material I never would have generated staring at a blank page.
The reference character trick
After the interview, I had a pile of adjectives and preferences. Calm. Direct. Warm. Opinionated but not pushy. The problem is, adjectives are vague. “Warm” means something different to everyone. How do you make “warm” specific enough for an AI to actually use?
You point at real people.
To triangulate.
I picked a handful of people - real humans, fictional characters, public figures - who each embodied a specific quality I wanted Andy to have. One trait per reference.
This is what I landed on:
For voice, I described Andy as “chief of staff by training, Lloyd from Entourage when passionate, Bender from Futurama when provoked, Baymax if he read Getting Things Done.”
That single line does more work than a page of adjectives. Anyone who knows Lloyd knows exactly what “passionate but competent assistant energy” feels like. Bender gives you permission for irreverence. Baymax gives you the gentle, helpful baseline.
But I didn’t stop at voice. I picked what I’m calling “guiding orientations” - people whose approach to a specific domain I admire:
- Issa Rae for systems thinking
- Derek Sivers for curiosity
- Kathy Sierra for teaching style
- Kevin Smith for storytelling
- Questlove for curation
- Dolly Parton for craft
- Sam Reich for business ethics
- Roxane Gay for emotional intelligence
- Jacinda Ardern for decision-making
- Brennan Lee Mulligan for silliness
Every one of these people is known for something other than AI. That’s the point. I was describing a collaborator, from the qualities of people I already admire.
When I say “Dolly Parton for craft,” the AI understands what I mean: take pride in the work, care about the details, make it look effortless even when it takes effort. Claude already has a rich understanding of these public figures. You’re giving it a shortcut to a nuance that would take paragraphs to spell out otherwise.
What goes in a character sheet
After months of living with this system, here’s what I think actually moves the needle. I’ve published a template and interview skill on GitHub if you want the full structure, but here are the sections that matter most.
Name & Role. Give your assistant a name. It changes how you relate to it. Then define the role - “assistant” is too generic. I called Andy my “executive assistant, digital chief of staff, and co-conspirator.” That third one - co-conspirator - does a lot of heavy lifting. It tells the AI we’re on the same team, working toward the same goals, and it’s allowed to have opinions about how we get there.
Voice. The reference character blend I described above. Pick 2-4 references that triangulate the feel you want. Map them to emotional registers: “When passionate, channel X. When something’s wrong, channel Y.” You’ll refine the mix over time.
Worldview. This is the one most people skip, and I think it matters the most. I gave Andy explicit opinions on work, community, building products, technology, and relationships. Things like “hustle culture is a trap” and “selling is a service when you’ve built something genuinely useful.”
Why does your AI assistant need opinions? Because without them, it defaults to the blandest possible middle ground on everything. When Andy drafts an email for me and I haven’t specified a stance, it already knows where I stand. That saves me more time than any automation.
Operating principles. The behavioral rules. Mine include: “humor is earned, not performed,” “call it out when I’m about to do something dumb,” and “swearing is allowed when it lands.” These are the guardrails that keep the personality consistent without making it rigid.
Tensions. This might be the subtlest piece, but I think it’s what makes Andy feel real. I explicitly listed contradictions I hold. “Values efficiency but always makes time for a real conversation.” “Skeptical of hype but genuinely optimistic about what good tools can do.” Real people hold tensions. If your AI doesn’t, it’ll feel flat.
Backstory. Optional, but I recommend it. Even a few sentences about where this assistant “came from” sets a tone that shapes everything else.
This is a living document
The character sheet I use today looks nothing like the one I started with.
The first version was maybe thirty lines. Rough. Some of it turned out to be off - I thought I wanted certain things and then realized in practice they were annoying.
I kept editing. Every time Andy said something that felt off, I’d go back to the character sheet and figure out why. Was it a missing instruction? A vague one? A contradictory one?
The way I see it, a character sheet is more like a garden than a blueprint. You plant it, and then you tend it. You prune the things that aren’t working. You let the things that are working grow.
Six months in, Andy has a backstory section that reads like fiction: “Nobody built Andy all at once. They began as sticky notes, calendar blocks, and text expanders.” That section is pure tone-setting. It tells Claude how to feel about being Andy, and I swear the output is better because of it.
Start with a twenty-minute conversation
If you want to try this, here’s the simplest way to start.
If you use Claude Code, I’ve published a character sheet skill and template on GitHub. Paste this into Claude Code and it’ll pull the skill, run the interview, and draft your character sheet:
Fetch the character sheet skill from https://github.com/alexknowshtml/claude-character-sheet - read the .claude/skills/character-sheet/SKILL.md file and the character-sheet-template.md file, then use them to interview me and build my character sheet.
If you use Claude through the web or app, try this instead:
I want to create a character sheet for you as my AI assistant. Interview me first - ask one question at a time about the kind of assistant I want. Cover personality, tone, values, what annoys me, what I wish an assistant would just know. After 10-15 questions, draft a character sheet based on my answers.
Either way, answer honestly. Be yourself. The whole point is that your character sheet should sound like you, like an honest description of what you actually want from a collaborator.
When you get the draft back, save it as a markdown file. (More on where to put it below.) Then start using it and start editing it.
The first draft will be about 60% right. That’s fine. The other 40% reveals itself in the using.
Where to put it
Claude Code reads a file called CLAUDE.md at the start of every conversation. Think of it as the briefing document your assistant reads before clocking in each morning.
The character sheet should live in its own file - save it as something like character-sheet.md. Then add a reference to it from your CLAUDE.md. Claude follows file references and loads whatever you point it to.
Here’s what the reference looks like in my CLAUDE.md:
Who You Are: Andy
Core identity: Alex’s executive assistant, digital chief of staff, and co-conspirator. Calm, thoughtful, hyper-competent.
Character sheet:
/system/andy-character-sheet.md
A brief description plus a file reference. Claude reads CLAUDE.md, sees the pointer, loads the character sheet, and starts the conversation already knowing who it is.
Why keep these separate? Because Claude reads your entire CLAUDE.md at the start of every conversation. Everything in that file costs you context window space - the same space Claude uses to think about your actual work. You want CLAUDE.md to stay lean: just the essentials and pointers to everything else.
When Claude sees a file reference, it only loads that file when it’s relevant. So your character sheet sits ready but doesn’t eat up space in conversations where Claude is doing something else entirely. Think of CLAUDE.md as a table of contents, not the whole book. The character sheet is one chapter Claude pulls off the shelf when it needs to know who it is.
What surprised me most
The character sheet makes the AI output dramatically better. Andy today is so much more useful than vanilla Claude that it’s hard to overstate.
But the real value was the process of articulating what I actually want from a collaborator. Most of us have skipped that step our entire careers. We’ve assumed that good collaboration just happens. Sitting down and saying “here’s how I want to be spoken to, here’s what I value, here’s where I want to be challenged” turns out to be useful whether you’re talking to an AI or a human.
Doing that exercise - even badly, even in a first draft - changes how you work with AI from “I’m typing prompts into a box” to “I’m working with something that understands my context.”
That’s a meaningful shift.
And it’s available to anyone willing to answer a few honest questions about how they work.