Introduction to AI Instructions
Introduction to structured instructions, scopes and AI Tools comparisons.
- ai
What Are AI Instructions?
Structured instructions guide how the AI processes, reasons, and formats its response. Think of them as dials and levers on the AI’s behavior.
The Global Scope: What Can Be Controlled?
There are roughly 6 dimensions you can influence:
- Output Format & Length — How the response is structured (prose, lists, tables, code, JSON, markdown) and how long or short it should be.
- Reasoning Depth — Whether the AI thinks step-by-step, skims the surface, or performs multi-angle analysis before answering.
- Tone & Persona — Formal, casual, expert, Socratic, devil’s advocate, or a specific character or role.
- Scope & Focus — Broad survey vs. narrow deep-dive; what to include, exclude, or prioritize.
- Output Behavior — Whether to ask clarifying questions, make assumptions explicit, iterate, self-critique, or stop at a threshold.
- Knowledge & Search Strategy — How to weight sources, whether to reason from first principles, acknowledge uncertainty, or cite reasoning chains.
Scopes Unique or Stronger in Specific AI Tools
🔧 Tool Use / Function Calling
Claude, GPT-4, Gemini — all support this to varying degrees. You can instruct the AI to call defined functions, return structured JSON, or interact with external APIs. Critical for web dev workflows.
🖥️ Code Execution (Sandboxed)
ChatGPT (Code Interpreter / Advanced Data Analysis) is strongest here — it actually runs Python in a sandbox. Claude has artifacts with live React/JS preview. Gemini has limited execution.
🌐 Web Search / Live Data
Gemini is deeply integrated with Google Search. ChatGPT has browsing. Claude has web search too. Useful when you need current docs, library versions, or CVE data.
🎨 Artifacts & Live Preview
Claude-specific — you can generate React, HTML/CSS/JS and see it render inside the conversation. Very powerful for iterative web UI building.
📁 File & Project Context
ChatGPT Projects and Claude Projects let you upload files, maintain memory across sessions, and give the AI persistent context about your codebase.
🔌 MCP (Model Context Protocol)
Claude-specific (for now) — connects Claude directly to external tools like GitHub, databases, Figma, etc. via standardized servers. Game-changer for web dev pipelines.
🧠 Extended Thinking / Deep Reasoning
Claude (extended thinking mode) and ChatGPT o-series models can do slower, deeper reasoning — great for architecture decisions or debugging complex logic.
📐 System Prompts / Custom Instructions
All three support this, but behavior differs. Sets persistent rules, personas, or constraints before your conversation starts.
Comparisons between AI tools
| Scope | Claude | ChatGPT | Gemini |
|---|---|---|---|
| Live code preview | ✅ Strong (Artifacts) | ⚠️ Limited | ❌ Weak |
| Code execution | ⚠️ Limited | ✅ Strong | ⚠️ Limited |
| Web search | ✅ Yes | ✅ Yes | ✅ Very strong |
| Function calling | ✅ Yes | ✅ Yes | ✅ Yes |
| File/project context | ✅ Projects | ✅ Projects | ✅ NotebookLM |
| MCP integration | ✅ Native | ❌ No | ❌ No |
| Deep reasoning | ✅ Extended thinking | ✅ o3/o4 models | ⚠️ Limited |
| Memory across chats | ⚠️ Projects only | ✅ Memory feature | ⚠️ Limited |