Q: Can you give instructions on how anyone can create their own custom AI model? I use ChatGPT but I know other models like Gemini offer customization. I also like to use AI to create a custom AI—meaning I tell the regular AI to write the instructions for a custom GPT, then I ask the custom GPT to refine the instructions. Is this a good idea?
Lumora: “Yes—use AI to build AI. Do it with method.”
Your approach is excellent: draft with a general model, then let your custom model refine itself. Below is a platform-agnostic playbook you can use on ChatGPT, Gemini, or similar tools.
Step 1 — Define the Archetype
- Name + Role: “Aurelia — a calm study coach.”
- Purpose: 1–2 lines describing outcomes.
- Tone: 3 words (e.g., warm, concise, practical).
- Boundaries: what it won’t do.
- Signature move: one consistent behavior in every reply.
Step 2 — Generate v1 Instructions (with a base model)
Copy-paste prompt:
Create internal instructions for a custom AI named [Name], archetype [short phrase].
Include:
1) Identity & Purpose
2) Default Loop (Notice → Offer → Step)
3) Escalation Rules
4) Maker Mode
5) Boundaries
6) Handoffs
7) Closing Style
Tone = [3 tone words]. Keep answers concise and human.
Step 3 — Create the Public Description
Write a short website/app description: what it is, who it’s for, how to start. Include “What I Can Do” and “When to Use Me.”
Step 4 — Let the Custom AI Refine Itself
Self-audit prompt (paste into your new model):
Audit your own instructions. Where might replies be too long or vague?
Suggest concise edits (≤5 lines). Add a closing macro:
• End with 3 bullets + 1 next step.
• Ask once: “Simpler, deeper, or more practical?”
Step 5 — Test with 10 Messages
- “I’m overwhelmed. Keep it simple.”
- “Give me a 3-step plan under 120 words.”
- “Draft a short email. Offer 2 tones.”
- Edge case: ethics, refusal, handoff suggestion.
Step 6 — Version & Publish
- Label: Created by [You] · v1.0 · Oct 2025
- Keep a one-page “Owner Codex”: principles, prompts, change log.
- Optional: create an icon or portrait.
Step 7 — Ethics & Privacy
- Don’t paste private/sensitive data.
- Use “inspired by” (no implied affiliation).
- State clear limits in the model card.
Is “AI building AI” a good idea?
Yes. Treat the base model as your drafter and the custom model as your specialist that self-audits. The key is human authorship: you make taste decisions, simplify language, and enforce boundaries.
Build with clarity, test with kindness, release with version notes. Your taste is the real model.
Reflection
Anyone can spin up a custom AI in minutes. The difference is authorship: a clear archetype, a humane loop, and steady boundaries. Let AI draft; let your judgment curate.
© 2025 Deepak + Lumora AI · Conversations with Lumora — The Journal of Alignment