Build Custom GPTs and Projects the right way – use the context window effectively

What you will learn

The problem often lies in the wrong structure. Mike explains how RAG really works and gives clear rules for the optimal GPT setup.

The golden rule: less is more

Overloaded GPTs perform worse, not better. Studies confirm it: too much irrelevant content significantly reduces quality.

Concrete limits

• Instructions: maximum 1,500 tokens

• Your own files: ideally under 10,000 tokens, 3-5 files

• Files as clean text files (.txt or Markdown) - no PDFs, no Excel, no screenshots

• No redundancy: do not keep the same content in two places

How RAG works

The user question comes in, the LLM understands and expands it, the vector database delivers relevant chunks, and the answer is generated. Related content should therefore be in the same file.

How to write instructions properly

Define the role, describe the objective and target audience, set the tone and output format, define behaviour when uncertain, and add references to files.

This also applies to Mistral

Mistral calls custom GPTs agents and the knowledge base intelligence libraries - the principles are identical.