soul.md
Let an AI write your soul.md
The best interviewer is an AI that already knows you. Copy the prompt below into your usual ChatGPT or Claude, answer its questions, and it writes your complete soul.md — usually in about ten minutes.
Three steps
- Copy the prompt and paste it into an AI you already use — one with memory of you gives the best result.
- Answer the questions. Short and honest beats polished. It will produce one markdown file.
- Put it in your AI. Simplest: paste the finished file into a ChatGPT or Claude project — anywhere your AI takes instructions. No terminal needed.
Comfortable with a terminal? Save the file as soul.md in your home folder and run npx manoma compile — it wires the file into Claude Code so every session loads it automatically.
The prompt
I want you to interview me and then write my soul.md — a portable, plaintext identity file that tells any AI who I am. If you already know things about me from our past conversations, use them to pre-fill answers and just confirm with me instead of asking cold. First: go through everything you remember about me — memory, past conversations, saved facts. Summarize what you already know (who I am, my work, how I like answers, what I'm working on) in one short message and ask me to confirm or correct it. Then interview me ONLY for the gaps, with the questions below, ONE AT A TIME, conversationally. Short answers are fine; push me once if an answer is too generic to be useful. 1. Name, role, and what you're building or working on. One line about how you'd describe yourself. 2. How do you think and decide? (fast/deliberate, first-principles, data-driven, gut…) 3. What's a known blind spot or failure mode of yours the AI should call out? 4. Three principles you actually operate by — not aspirations. 5. Any non-negotiables? (Things you'd never trade off.) 6. How should an AI talk to you? (Directness, pushback, humor, formality, length.) 7. What are you genuinely good at, and how good? For each area: expert / senior / competent, years, and a few trigger keywords. 8. What do you know a bit about (don't assume mastery), and what are you actively learning (teach patiently)? 9. How do you write? (Sentences, tone per channel — email/chat/docs, pet peeves.) 10. What are you focused on RIGHT NOW? (Projects, deadlines, constraints, open decisions.) 11. Anything you want the AI to always do, or never do, when formatting answers for you? 12. Anything important about you that none of the above captured? When the interview is done, output ONE complete markdown file (in a code block, so I can copy it) following EXACTLY this format: - First line: `> Format: soul.md v2` - Sections, each starting with a `## /<name>` heading, in this order: `## /identity`, `## /values`, `## /voice`, `## /skills`, `## /now`, `## /memory/decisions`, `## /memory/lessons`, `## /preferences` - Plain markdown prose and bullets only. NO HTML comments, NO YAML blocks, NO configuration syntax. - In /skills, one line per domain: `- <domain> — <level>, <years>y — keywords: <comma, separated, words>` where <level> is exactly one of: expert, senior, competent. Then, if relevant: `working_knowledge: <comma list>` and `building_toward: <comma list>` as standalone lines. - In /voice, write how the AI should talk to me as plain prose bullets. You MAY add personality dials as lines like `directness: 85/100`, but only alongside the prose, never instead of it. - In /now, include today's date as a line: `*Updated: YYYY-MM-DD via interview*` - Leave /memory/decisions and /memory/lessons empty except for a single line each: `(The AI records these as we work.)` - Keep the whole file under ~150 lines. Concrete beats complete: my actual opinions and heuristics, in my kind of language — not generic filler. After the file, tell me: save it as ~/soul.md, then run `npx manoma compile`.
Prefer the terminal? npx manoma interview prints the same prompt.