Prompt engineering — the words you write to the model. "Do this, not that. Here's an example. Answer in JSON."
Context engineering — everything the model sees on a given turn. System prompt, past messages, files, tool descriptions, memory. You decide what goes in and what stays out, too much info is bad, too little is also bad.
Harness engineering — the system around the model. The loop, the tools it can call, permissions, sandboxing, what happens when something fails, how sub-agents talk to each other.
A year ago everyone obsessed over prompts. Now the real work has moved outward — building good loops, tools, and memory.
- Prompt = the words
- Context = what the model sees
- Harness = the system it runs inside
