The Cornetto Ice Cream Cone Framework for Prompting

A prompting framework called the Cornetto Ice Cream Cone Framework, also referred to as the "Four C" framework, organizes prompts for medium-to-large agent tasks around four elements: context ("the why"), constraints ("the how, and the how not"), and a control loop that is deliberately kept independent of the other components, typically implemented with subagents using a fresh context window or static checks such as red-green tests. The framework's author recommends embedding a short version in the system prompt, noting that research tasks and one-off questions or simple parsing jobs do not fall under the medium-to-big task category.

The Cornetto Ice Cream Cone Framework for Prompting Four C’s for clarity on complex tasks Here’s a simple framework I use for improving the performance of your fav agent on semi- complex tasks: I think it’s all mostly obvious how the four C’s improve your prompt, which means I won’t explain each part of the framework. The only subtle part is the control loop. Designing the control loop to be more “independent” is supremely useful; in practice it means the control loop doesn’t know about the other three C’s. I usually use AI agents with a fresh context window for this. But the independent design principle can materialize in all sorts of things: static tests in programming, a screenshot of the website the agent is working on, or, if you want, a human in the loop. Maybe, just maybe, you can let the control loop know about some of the constraints. I never go overboard with my prompts, and I think the framework is simple enough to scale well with model intelligence. You can also call it the “Four C” framework. If you think you will forget about this anyway, put something like this in the system prompt: ask for these c's for medium to big tasks Medium to big tasks are tasks that are not one off, aren't a simple question or something like filling out a form or parsing a pdf, it's building something new or synthesizing multiple things. Research doesn't fall under this. context the why constraints the how, and the how not control aka, the loop the controlling, importantly, the control doesn't know about the context and constraints, usually subagents or more static control like red-green tests.