Published: 2026-05-21

Stop Prompting, Start Questioning: The Senior Partner Method for AI Agents

Chapters / key moments (click to jump — plays here on the page)

Nate B Jones argues that prompt engineering is now table stakes — frontier models like Opus 4.7 and GPT-5.5 are roughly 100x more capable for agentic work than models from six months ago, yet most people still interact with them like junior assistants. His "AI Question Method" reframes the interaction: instead of specifying each task, you convey your perspective and thesis, then ask questions that open up the problem scope and invite the agent to bring its full capability to bear as a senior partner.

Source video

"You're Not Bad at Prompting. You're Bad at Defining the Work." by Nate B Jones — Watch on YouTube →

Key Takeaways

Prompt engineering is now table stakes — you don't get credit for it. The next skill layer is learning to work with AI as a senior partner, not a task executor.
Frontier models (Opus 4.7, GPT-5.5) are roughly 100x more capable for agentic work than models from six to eight months ago — our prompting habits haven't kept pace.
The AI Question Method: frame your interaction as a series of questions that define scope and invite exploration, rather than a task list that constrains the agent.
Always convey your perspective and thesis — vague requests produce generic output even with powerful models. Give the agent your angle before asking it to investigate.
This approach applies to heavy knowledge work in Claude Code, Claude Cowork, and CodeEx. Defined agentic pipelines (invoices, tickets) are a separate category and don't need it.
Since frontier models now have persistent memory, you can tell them to adapt to question-based interaction and they will carry that preference forward.

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Reproducible steps

State your thesis, then set the edges
Give the agent the centre of your intent and your hunch: 'I think attribution is broken because Google organic isn't bucketed correctly; investigate with that in mind, I might be wrong.' Then state hard exclusions: 'we spent 15 minutes of that meeting on an unrelated project, leave it out of the report.'
Ask several hard questions and let it synthesise
Instead of dictating structure, pose the difficult sub-questions you haven't answered yourself. His PR/FAQ example asked how to make the experience accessible to people who have never used 3D, and how to convey the hardware/software relationship in both the press release and the internal FAQ.
Name the data and invite pushback
Collect formal and informal inputs into one working-context folder (he has Codex copy docs, PRDs, support tickets, call transcripts and analytics into it). Then name those artifacts in your question, give your thesis, and say explicitly: you don't have to agree, come back with the most explanatory thesis across all the data.

Gotchas

Aimed at frontier models on paid plans doing knowledge work (Opus 4.7, GPT-5.5 era). He says it won't work well on free tiers or older models.
A question that names only one file can steer the agent into that file alone. Name the breadth you want covered.
Stating an opinion without inviting disagreement tends to get your opinion mirrored back.