Prompt engineering advice has evolved, prompting has moved from “tell the model every step” to “define the outcome, constraints, tools, evidence rules, and stopping conditions”. Now, as models become more capable and accept new modes (image/video), prompting advice would continue to evolve.
Prompt engineering as a design tool
Early prompt improvements would have looked somewhat like this:
Early power users recorded how to write good prompts. You can read a great example here:
- anatomy of a great prompt by Greg Brockman
- The prompt engineering guide by Nate
- Lenny's advice on prompt writing
Frontier Labs - Google, OpenAI and Anthropic now have thier own advice on how to write good prompts. Compare it across models to see if what we mentioned in the first sentence is true.
Prompt engineering > context engineering
With more context, traditional prompt engineering is now dead. Context engineering is: what information, tools, state, memory, examples, files, schemas, screenshots, and constraints does the model need to perform the task well? It defines the task.
I too find myself giving models even more context than ever before during my agentic engineering session.
Next frontier of prompting?
As models evolve and become more capable, we can expect prompt engineering advice to change. We saw this recently with the release of the new model by Thinking Machines.