Prompt engineering was all the rage in the ancient historical period known as November of 2025.
I've got something to tell you though. The models got better. The system prompts. The tool calling. The context window management. Now, we are really in an era where once you've surpassed a certain understanding of what the models are capable of, and how they assume certain modalities of operation, as long as they are kicked in the correct latent vector space direction you'll get absolutely fantastic outputs.
Let me explain the concept with a concrete example. This Australia Day long weekend, I told my wife I would be attempting to recoup the costs of the $200 USD Claude Code subscription by targeting a once off $200 of revenue. Bold, I know.
So she told me some of her pain points in her day, one of which stuck out. Grandma - currently staying with us is getting a bit bored as a non-english speaking elderly, non-driving (also not entirely literate even in Vietnamese) stuck in the panacea we call the Suburbs of Greater Brisbane. I began by jotting the idea into MicroSoft To Do. With all my faculties I managed to crop a screenshot of the layout of the task (Green) and used voice mode to articulate the remainder (Blue).
This was pasted into Clopus .45 and shot off into vector space for all B200's in existence to rumble on:
The result is the thing we care about.
In days of yesteryear this would've probably gone off the rails, or a few years back required an entire pack of university students to come up with such a project proposal.
You might be wondering, why am I sharing this idea? I might have a go at building it. I've gotta get to $200 MRR after all. But honestly, if you dear reader decide to build it - I'll be happy to become a paying customer anyway.
Your prompting correctness matters less than ever. What matters, is in fact execution. The ability to chain together outputs. To orchestrate the agents and let them rip at what they're good at.
Post SaaS, not prompts.