A decision framework for when to build a custom AI

Executive overview

Most users stay in basic chat and miss 80–90% of AI's potential. Custom GPTs and GPT projects unlock specialised, persistent, high-quality outputs for repeated tasks.

Three questions determine whether to build: Is the task repetitive? Does quality matter? Who needs access?

The gap between casual and advanced AI users is almost entirely explained by custom AI usage.

The three-question decision framework

• Is this task done more than once — daily, weekly, monthly?
• Does quality matter enough to justify setup effort?
• Is it for you alone, or shared with a team?

GPT projects vs custom GPTs

• GPT projects: personal use only; more features — persistent memory, deep research, agent mode
• Custom GPTs: shareable with a team or publicly via link
• Deep research produces McKinsey-level 20-page reports after ~30 minutes of autonomous research
• Agent mode lets the AI browse and complete tasks autonomously (beta)
• If sharing is not required, projects are the better default

Building a strong system prompt

• Use ChatGPT's GPT-5 prompt optimizer or Anthropic's console prompt generator
• Give the AI a basic intent; the optimizer rewrites it for higher-quality outputs
• Upload relevant files (PDFs, databases, text exports) as the AI's knowledge base

Example 1: YouTube title creator

• Feeds video notes into an AI trained on a database of proven title frameworks (Creator Hooks newsletter archive)
• Returns 3–6 optimised titles per session
• Maintains a persistent file of already-used titles to avoid repetition

Example 2: Tax specialist

• Start with an unmodified AI to conduct a structured interview — one question at a time — about business and finances
• Feed the interview summary into a high-reasoning model (GPT-5 thinking or Opus) to generate a tailored system prompt
• Bake that prompt into a GPT project for ongoing tax advice
• 4 out of 5 AI-informed recommendations were validated by a human tax specialist

Example 3: LinkedIn post writer

• Writer drafts a first post; the AI rewrites it using one of five proven LinkedIn formats (sourced from a Matt Gray PDF)
• Format selection is automatic based on topic
• Keeps posts consistent with what currently performs well on the platform

Other categories worth building for

• Contextual research assistants
• Sparring partners for intellectual debate
• Specialist advisors (legal, finance, health)
• Coaches (fitness, productivity)
• Therapists aligned to specific philosophies
• Data cleaning pipelines for recurring file formats