1 day ago
Developers are flipping the usual advice about AI. Instead of chasing the perfect prompt, they let the model interview them first. The technique, called reverse prompting, is simple: dump everything about your goals into the AI, then ask it what else it needs to know before it suggests a plan. It is spreading alongside a second idea, graph engineering, where people wire up networks of agents that check and improve each other's work through set workflows. The common thread is that structure and coordination now get more out of a strong model than clever wording does.
Reverse prompting hands the questions back to the AI. You give it context, then ask it to lead. A typical version looks like this: brain-dump your career, goals, and constraints, then send, "Based on what you know about me and my goals, what more can I tell you so you can help me get there faster?" Follow that with, "What can you do for me right now to move us closer to those goals?" The point is that a capable model often knows what to ask better than you know what to request. You stop guessing at the ideal instruction and let it pull the details out of you. People who try it tend to report a longer, more useful task list than they walked in with.
Graph engineering takes the same idea and scales it past one chat window. Rather than one model doing everything, you build a small network of agents with defined roles, and they verify and refine each other's output. One drafts, another reviews, another checks the result against the goal. The structure is what keeps the system honest, so it can run without someone watching every step.
Google has put out a free course on building agentic systems that walks through these ideas from the basics up to large setups, and it runs about an hour. The through-line is blunt: solid structure is what makes AI reliable. Get the environment and the handoffs right and you need far less babysitting.
The bigger takeaway is a shift in where the skill lives. For a while, prompt wording was treated as the whole game. Now the edge is moving to how you set up the work: what the agent knows, how agents coordinate, and how their output gets checked. As the underlying models get stronger, the environment around them is doing more of the heavy lifting than any single prompt can.
