Every time you send a message to ChatGPT, Claude, or any LLM, it forgets you exist. No memory of your last conversation. No context from before. Nothing. The only way it “knows” who you are is if you re-introduce yourself. Every. Single. Time. That is not a bug. That is the problem context engineering solves.
Most people blame the model when AI gives garbage outputs. Wrong diagnosis. The model is not the problem. The context is. Everything an LLM needs to answer your question must be sent WITH your question. Understanding this one thing will change how you work with AI.
Context has 4 building blocks. Think of each as a dial you can tune.
1/ Memory
Your conversation history and preferences. The catch? LLMs have limits. Long conversations get summarized, and that is when “drift” happens. You start repeating yourself. The AI forgets your rules. That is not the model getting dumber. That is imperfect memory management.
2/ Files
Documents, screenshots, PDFs fed directly into the context. Hot tip: drop a screenshot into ChatGPT and ask it to build a prompt from it. Works shockingly well.
3/ RAG (Retrieval Augmented Generation)
Before your question hits the LLM, smart systems quietly search a knowledge base and stuff relevant results into the context alongside it. That is why internal AI chatbots can answer questions about your specific products. Not magic. Retrieval.
4/ Tools
The LLM does not execute tools. It tells your system WHAT to run. Your system runs it, stuffs the result back into the context, and THEN the model answers. Tool calling is the AI directing traffic, not doing the work.
Most people treat prompts like a black box. Throw stuff in, hope for the best.
But when you see memory, files, tools, and prompt structure as separate levers, you stop guessing and start engineering.
The 6 prompt components that actually matter:
✓ Role: “You are an expert in...”
✓ Personality: tone and style
✓ Request: the actual task
✓ Format: be explicit (bullets, JSON, table)
✓ Examples: two good, two bad works wonders
✓ Constraints: “never do X”
Most people only use the Request and skip the other five.
Here is the bigger picture.
Context engineering is not just for developers building agentic systems. It is a thinking skill.
When you learn to structure information so a machine can reason over it accurately, you get better at structuring information for humans too. Clearer writing. Clearer thinking. Clearer communication.
Context is not a setting you configure once. Context is everything you send.
Get that right and the model almost does not matter.
Are you still treating AI like a magic black box, or are you already thinking about context?
Drop a comment. I am curious where people are at with this.
Originally posted on my Substack
Top comments (0)