Your prompts are good now. Role, context, goal, audience, constraints, examples, the lot. And then you ask for something genuinely large, a full training curriculum, a business plan, a launch strategy, and what comes back reads like the work of somebody who started strong and quietly lost interest around the halfway mark.
Nobody lost interest. You asked one question that contained six questions, and the model answered all six in a single pass, with one pool of attention split six ways.
This is worth checking on your own output. Take the last very large thing you asked for and read it from the bottom up. Section one is specific, structured, full of the concrete detail you wanted. Section six is a list of headings with a sentence underneath each one. Nothing failed. There was simply less left to spend by the time the model arrived there.
A large task is not one prompt. It is a pipeline, and every stage should receive the finished work of the stage before it.
The stages are already there
You do not have to invent the structure. Anything big enough to disappoint you in one prompt comes pre-divided:
- Stage 1 produces the skeleton. An outline, a section list, a structure.
- Stage 2 fleshes out one part. One only, in full depth.
- Stage 3 continues through the remaining parts, holding the format.
- Stage 4 checks the whole thing for consistency, gaps and transitions.
Each prompt in that sequence is smaller, narrower and better briefed than the mega-prompt it replaced, and each one produces better output for exactly that reason. This is not a trick for people who like process. It is what sits underneath production AI pipelines, and it is how AI-native products get consistent output at a scale no single prompt could hold.