Build AI Specialists with Projects and Gems Across ChatGPT, Claude, Gemini

Executive overview

Most people use AI reactively — one-off chats that require re-explaining context every session. The higher-leverage approach is building AI specialists: persistent, preconfigured agents that accept minimal input and return consistent, high-quality output. All three major platforms (ChatGPT, Claude, Gemini) support this via Projects or Gems, each containing two components: system instructions and knowledge files. The WWH framework (What, Why, How) structures the base prompt, which is then run through a built-in prompt optimizer before deployment.

Build once, use forever — then democratize it to your team.

What makes up an AI specialist

• ChatGPT and Claude call them Projects; Gemini calls them Gems
• Two components: system instructions (your prompt) and files (context documents)
• Files can be PDFs, Word docs, CSVs, text files — avoid large multi-tab Excel sheets; export the relevant tab as CSV instead
• Instructions define role, task boundaries, and output format
• Projects must be visited; Claude Skills can be called into any conversation

The WWH prompt framework

• What: the specific task, described in detail
• Why: the purpose — lets the AI infer unstated requirements
• How: optional constraints on style, format, or output type
• Run the drafted prompt through OpenAI's or Claude's built-in prompt optimizer before deploying
• Including Anthropic's research snippet inside a prompt improves AI research quality 2–3x

Three example specialists

• Tax advisor (ChatGPT): fed with business context files, scoped strictly to US tax strategy, instructed to research only current strategies
• Document formatter (Claude): combines a Project with a Skill that reverse-engineers a template's fonts, colors, and spacing to produce pixel-perfect SOPs and contracts
• YouTube title generator (Gemini): merges a framework database of high-view titles with a client-conversion reference file to balance reach and revenue

Sharing and model selection

• Projects are shareable across teams; ChatGPT Teams supports permission tiers (view vs. edit)
• For recurring high-quality tasks, use the most capable model available: GPT-5.2 (extended reasoning), Claude Opus 4.5 (extended thinking), Gemini 3 Pro
• A/B test the same specialist across all three platforms over time to find consistent winners per niche

When to build a specialist

• Ask three questions: Does this need consistent high-quality output? Do I do it 3+ times per week or month? Does my team also need access?
• Two or more "yes" answers means the task warrants a dedicated AI specialist
• Not every task qualifies — one-off or low-frequency asks are better handled in standard chat