AI assistants

Use this page when the main problem is how an assistant product behaves: evidence rules, memory boundaries, or browsing inside ChatGPT, Claude, Gemini, or a similar assistant.

Use this page when the main issue is how ChatGPT, Claude, Gemini, or another assistant behaves: which sources it may use, how memory affects replies, or when web browsing is actually needed.

Prerequisites

Before you start

• Know which assistant product you are using and what information it can access.
• Decide whether the task should use only uploaded material, cited public sources, or no external retrieval at all.
• Decide whether web browsing is really necessary before turning it on.

Scope

When to use AI assistants

• Use this page when the work happens inside a chat assistant product, not inside an agent loop or an API integration.
• Use it when you need to control where the assistant may get facts.
• Use it when you need to manage how memory, files, or prior chat context affect the answer.
• Use it when you need web browsing only for current or niche public information.

Use another topic when the main problem is elsewhere:

• Prompt engineering — The main problem is how to write the prompt, not how the assistant product behaves.
• Agents and workflows — The main problem is a multi-step system with routing, orchestration, tools, or memory across steps.
• API integrations and tools — The main problem is connecting the model to tools, APIs, web access, or another system.

Continue with the right resource

Choose the resource that matches the next action, not just the content type.

Guides

Use this when the assistant already has the right source rules and you now need current information from the web.

• Request web browsing in an assistant Use this when live public information is really required and you want a controlled browsing workflow.

Policies

Use these when the assistant must follow fixed rules about sources, memory, and web use.

• Allow facts only from uploaded files Use this when the assistant must stay inside the files you provided.
• Require cited public sources for facts Use this when factual claims must come from cited public sources.
• Require web checks for current or niche claims Use this when the answer depends on recent or specialized public information.

Prompts

Use these reusable prompts when you want to control sources, priorities, or web use inside an assistant product.

• System prompt: use only uploaded files for facts Use this when the assistant must not bring in facts from anywhere else.
• System prompt: require cited public sources Use this when factual claims must come from cited public sources.
• System prompt: keep instruction priority and source rules clear Use this when you want tighter control over which instructions come first and which sources are allowed.
• Workflow prompt: request web browsing Use this when you want a controlled browsing request inside the assistant.
• Workflow prompt: run a web check Use this when the task depends on current public information and needs explicit checking steps.
• Workflow prompt: return citations in a fixed format Use this when the final answer must present citations in a fixed structure.

Reference

Use these when you want the background behind how assistant requests and connectors work.

• Reference: how ChatGPT builds the final request Background model for how assistant-side context is assembled before generation.
• Reference: how ChatGPT connectors change between basic mode and agent mode Use this to understand how connector behavior changes across operating modes.

Articles

Use these when you need a deeper explanation of memory behavior and assistant-specific failure modes.

• Read the LLM Memory Boundary Model Long-form explanation of memory scope, carryover, and boundary failures.
• Read Sycophancy in LLM Assistants Long-form explanation of agreement bias in assistant products.