ELI5: System Prompts vs Models

The Model vs. The System Prompt

The model is the engine. Claude, GPT-4, Gemini — raw horsepower. How fast it thinks, how deep it reasons. Fixed. Under the hood.

The system prompt is how you drive. Instructions, rules, personality, guardrails. Same engine. Completely different ride.

The model sets the ceiling. The system prompt decides if you ever touch it.

What the research shows

Six AI coding agents. Same job — write code. All powerful models inside. Wildly different system prompts. Same engine. Very different drivers.

The extreme end

Cursor burns over a third of its system prompt on personality and steering alone. Kimi CLI? Almost nothing. Shortest prompt. Zero workflow guidance.

Same destination. One driver has a route, a map, and a rule for every junction. The other just pulls out of the driveway.

Scar tissue

This is the part I keep thinking about.

System prompts are full of patches — repeated instructions, ALL CAPS commands, words like MUST, NEVER, ALWAYS.

Those aren't just rules. They're crash reports. Someone drove this engine before you. It veered. They wrote it down.

The scars tell you exactly where the engine misbehaves.

If you're building agents

Picking the model is step one. Not the whole game. Can a great driver in a mid-range car beat a bad driver in a Ferrari? Often, yes. Your system prompt is your driving. Learn to drive well.

Source article: https://www.dbreunig.com/2026/02/10/system-prompts-define-the-agent-as-much-as-the-model.html

Tool worth exploring: https://nilenso.github.io/context-viewer/