Full Post w/ Prompt Pack: https://natesnewsletter.substack.com/p/ai-agents-better-communicator?r=1z4sm5&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true __________________________________ What's really happening with prompt engineering now that AI agents are 100x more powerful than they were six months ago? The common story is that prompt engineering is dead and you can just ask AI for what you want — but the reality is more complicated. In this video, I share the inside scoop on the shift from prompting to questioning AI agents: • Why prompt engineering is now table stakes, not a skill • How to treat frontier LLMs like a senior partner • What the AI Question Method looks like in practice • Where most knowledge workers are still stuck in 2025 habits The risk is that anyone running heavy knowledge work with weak questions gets shallow output from powerful agents, but the operators who learn to ask sharp, layered questions unlock real leverage. Chapters: 00:00 Prompt engineering is now table stakes 01:24 Why agents are 100x more powerful in 2026 02:48 Introducing the AI Question Method 04:05 AI as senior partner, not junior teammate 06:20 Why most people are still prompting like it's 2025 08:10 Defining agents vs. agentic pipelines 10:05 Principle 1: The flashlight intent 12:30 Conveying perspective and edges in your questions 14:45 Principle 2: Asking what good looks like 16:20 The PRFAQ example with Prime Video 19:10 Principle 3: Wrestling with data and opinions 21:30 The MRR product-led growth example 23:45 Why memory and quick-start guides matter Subscribe for daily AI strategy and news. For deeper playbooks and analysis: https://natesnewsletter.substack.com/ Listen to this video as a podcast. - Spotify: https://open.spotify.com/show/0gkFdjd1wptEKJKLu9LbZ4 - Apple Podcasts: https://podcasts.apple.com/us/podcast/ai-news-strategy-daily-with-nate-b-jones/id1877109372
ADVERTISEMENT
I'm on sonnet thinking with planning then opus for execution, context management and having a fit for purpose harness has been my most cost effective approach for me. You need to learn to plan first execute last, artifacts are key along the way. Throttling has reorganized my strategy and reduced my dependence on AI, after every prompt/session it's a must that everything I have developed (AI assisted) should be able to be continued by me (AI unassisted) , making an AI blackout a non-event for my workflow. Is it fast? No, but is it measured yes. Artificial intelligence at best is an augmentation to human intelligence not a replacement.
shit changes so much i dont believe anyone has an actual job within these realms for more than a week
A good manager may ask those questions but a good employee will ask your intent, why do you want to do this... That's how are good employee achieves good outcomes from a vague manager..
Communication is the most valuable skill in the future.
I use Claude Code and Codex to manage my Hermes… senior partners, managing the junior holding a hand grenade. My communication patterns with AI has improved my communication with my kids😂😂😂
Throughout this whole video, I kept thinking that we’re in a weird golden period where these things have become peers. If the shift is from junior to senior these last six months, where does that put us in a year, 2 years, 5? Things are getting weird, folks.
Thank you, Nate 🎉
With Opus 4.7, I've begun having 'conversations' with it as if it were just another person. I almost (but not quite :P) forget that I'm conversing with AI.
Do you have a discussion community? Looking for a place to collaborate with others and learn about how to set up my AI stack to accomplish my code based projects most efficiently and get the most out of claude
/grill-me for the win!
for vibe-coders here - Nate is talking about "open-ended" specifically. so you are ok with your 2025 prompts
When I need to do a deep dive, my approach has been to approach the conversation Socratically where I ask questions which lead in the direction I'm thinking.
You are one of the rare ai content producers who obviously and genuinely knows what they are talking about and are clearly experienced and practiced that I follow. Never thought I would have this thought but I just said “man I wish this guy had a podcast” LOL
9:30-9:59 the issue i hit when i do this is confirmation bias issue. Where the model just finds a way to agree with whatever perspective i have. How do you prevent that?
Knowledge workers may be discovering human-Ai partner paired co-working, but any vibe coder who hit the wall and realized they need to understand what they are building, has been doing this already.
spark
Dude loves endless talks. Looks like a clown.
I’ve been doing this for a while now and it definitely got way better with the recent models. The way I usually try to explain is like discussing based development. I just talk with my agent until I’m confident we both understand the problem very well, then we talk it out until im confident we have a solution, and he builds it. No fancy prompts, no powerful harness. Just the vanilla model and some efficient knowledge base.
Its the workflow; the model (the tool box), prompt (the guard pipeline), and harness (the protection to stay on topic and what not to do) is not lost. It's still the prompt and tools used thats doing the work in AI. You just need the flow to matter the most in the layers of the action. So the context just need more a structural architect still (hierarchy); but its just less now. Prompts not dead, the baseline just got wider. In complexity its still needed depending on the task needed. [RAG got better now if anything.]
I never noticed you're local dude. Crazy!