Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Opus 4.7 and OpenAI 5.5 Made Your Prompting Style Obsolete.
TL;DR
Prompt engineering is now table stakes, not differentiation — Nate B Jones says the old 2025-style obsession with precise prompting is “dead” because newer models like Opus 4.7 and OpenAI 5.5 are dramatically better, so the real skill is what comes after basic prompting.
The big mental-model shift is from junior partner to senior partner — instead of micromanaging AI step-by-step like last year, he argues you now get the most value by working with frontier models as if they’re senior teammates who can push back, synthesize, and reason across messy knowledge work.
His replacement for prompting is the “AI question method” — rather than issuing a task, you frame a problem with questions, a thesis, and boundaries, like his Amazon manager who handed over CSVs and Excel files, explained the story she wanted for a deck and doc, and guided the work through questions.
Good AI questions need a ‘flashlight center’ plus hard edges — his key advice is to give the model a directional point of view (“our marketing attribution may be broken because Google organic is misbucketed”) while also excluding irrelevant material, like cutting 15 minutes of unrelated meeting chatter from a report.
The strongest workflows ask AI to wrestle with both explicit files and implicit opinions — Nate’s example is loading customer transcripts, support tickets, PRDs, launch docs, and MRR analytics into Codex or Claude/Co-work, then asking for the most elegant thesis across all of it, even if the AI disagrees with his product-led-growth hypothesis.
This only really works on cutting-edge, high-context models — he repeatedly says the approach is for heavy knowledge work in tools like Claude Code, Codex, and Co-work with frontier models, not for older models, free-tier accounts, or tightly defined agentic pipelines like invoice or support-ticket automation.
The Breakdown
Why Nate says prompt engineering is “dead”
He opens with a provocation: prompt engineering was the right conversation for 2025, but now it’s just table stakes. The eye roll he keeps seeing is real — people think they can “just ask AI for what they want” — but his point is that this only works when you already know exactly what you want, which breaks down in complex agentic work.
Opus 4.7 and OpenAI 5.5 changed the interaction model
The catalyst, he says, is the recent jump from models like Opus 4.7 and OpenAI 5.5, which he calls “100 times more powerful” than agents from six to eight months ago. They call tools better, pull data better, and sustain longer work, so the mismatch now is that our prompting habits have not improved at anything close to the same rate.
The manager story that reframes AI as a senior partner
Nate’s core analogy comes from an old marketing job: a great manager at Amazon would hand him CSVs and Excel files, explain the attribution problem, say the output needed to become both a deck and a clear doc, and then guide him with questions instead of prescribing every move. That, he says, is exactly how you should talk to AI now — not like a junior employee who needs hyper-specific instructions, but like a senior partner you can think with.
What kind of “agent” he actually means
He pauses to define terms because “agent” is overloaded. He’s not talking about buttoned-up workflows for invoices or customer service tickets; he means deep knowledge work in tools like Claude Code, Codex, and Co-work, where the task is ambiguous, custom, and high leverage.
Principle 1: Give the AI a thesis, not just a task
His first rule is that your questions need a point of view — the “center of a flashlight beam” — plus boundaries. His examples range from asking about the Mona Lisa through the lens of Leonardo da Vinci’s later life to telling AI that a marketing attribution issue may stem from Google organic being bucketed wrong, while also explicitly excluding irrelevant meeting sections from a final report.
Principle 2: Ask questions that define what “good” looks like
The second move is asking the AI to reason about outcomes that are hard to reduce to a checklist or eval. He uses a fictional Amazon-style PR FAQ for Prime Video’s absurdly vivid example — 3D World Cup players kicking a ball around your living room — to show how you can ask the model to figure out customer accessibility, hardware-software interplay, and emotional messaging without overprescribing the answer.
Principle 3: Make it engage the whole corpus, not just one file
His final principle is about getting AI to wrestle with all the inputs: formal docs, transcripts, support tickets, analytics, and your own thesis layered over them. In his product example, he has Codex organize voice-of-customer data, PRDs, launch docs, and MRR analytics into one folder, then asks AI to test his belief that product-led growth is broken and return the cleanest explanatory thesis — even if that means proving him wrong.
The real takeaway: stop “prompting,” start managing
He ends by planting a flag: we’ve moved beyond the old prompt-engineering mindset, even if basic prompting still matters. The future skill is asking sharp, structured questions that let powerful models explore, synthesize, and challenge you the way a strong senior teammate would.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
