Nate B. Jones's Upgraded Expert Prompt: Which Part Works
By Matt Cullerton, founder of Mavric and Apso
Nate B. Jones's "Expert Prompt, Upgraded For 2026" stacks four instructions: act like an expert, go back to root causes and restate the problem, propose the simplest solution, and name the tools and skills needed. Published research finds the "act like an expert" line does little for accuracy, while the root-cause step matches a technique with measured gains. One line in the prompt, asking the model to "go back and look" at its training data, describes something no model can do.
What is the upgraded expert prompt?
In a TikTok from @nate.b.jones, Jones argues that people ask AI for help outside their own expertise, so the weak point is the person asking. His fix is a prompt that tells the model:
- "You need to act like an expert in this domain."
- "Go back to root causes, fully identify the problem, even if I haven't expressed it correctly."
- Propose the simplest solution.
- If "tools or skills" are involved, "find the simplest, clearest possible configuration of skills and tools and say, this is what I need to solve this problem for you."
He calls it an update to "the old, you're an expert prompt," and says "this is now for 2026 and it works better." The video is a single talking-head shot with no chat window, model output or comparison on screen.
Does "act as an expert" improve AI answers?
Not for accuracy, according to the best-known study. Mingqian Zheng, Jiaxin Pei, Lajanugen Logeswaran, Moontae Lee and David Jurgens tested this in "When 'A Helpful Assistant' Is Not Really Helpful," published in Findings of EMNLP 2024. They used 162 roles across 4 LLM families and 2,410 factual questions and found that "adding personas in system prompts does not improve model performance across a range of questions compared to the control setting where no persona is added."
Anthropic's prompting best practices make a narrower claim: "Setting a role in the system prompt focuses Claude's behavior and tone for your use case." That covers behavior and tone. It makes no promise about factual accuracy.
Why the root-cause instruction is the useful part
The second instruction has a published equivalent with numbers. "Take a Step Back" by Huaixiu Steven Zheng and colleagues, submitted to arXiv on 9 October 2023 and accepted at ICLR 2024, describes Step-Back Prompting as "a simple prompting technique that enables LLMs to do abstractions to derive high-level concepts and first principles from instances containing specific details."
On PaLM-2L the paper reports gains of 7 percent on MMLU Physics, 11 percent on MMLU Chemistry, 27 percent on TimeQA and 7 percent on MuSiQue. It also ran experiments on GPT-4 and Llama2-70B.
Both ask the model to drop the user's framing and rebuild the general problem underneath. Jones says as much: "you're basically telling the AI, don't get locked into what I said because I'm not the expert." The title card credits the persona line, but the evidence supports the step-back instruction.
Can an AI look up its training data?
No. Late in the video Jones says "go back, you have training data, go back and look and see what a proper expert would do with this." A deployed model has fixed weights and no index of the text it was trained on. Asking it to look changes which words it finds likely. It does not start a lookup. Retrieval needs an attached tool such as web search.
Anthropic's documentation says the opposite for high-stakes answers. Its medication dosing guidance has Claude give a child's dose "from the product's label, not from memory," and says that without search, "Claude doesn't give a number from memory and instead points the person to the label's dosing chart, a pharmacist, or the child's doctor."
This matters for the examples in the video, such as someone "having trouble with real estate listings." Real estate figures are local and change often. A prompt that leans on training data returns a confident answer with no way to tell current numbers from stale ones. Pasting in a source document or turning on search fixes that.
What "skills" means in 2026
Jones uses "skills" in the everyday sense. Since Anthropic's 16 October 2025 engineering post, the word also names a specific format: "organized folders of instructions, scripts, and resources that agents can discover and load dynamically." A skill is a directory with a SKILL.md file whose YAML front matter includes name and description. agentskills.io describes it as an open standard "originally developed by Anthropic."
A plain chat session cannot install or enable a skill or tool. The prompt's answer on that point is a list you then act on yourself.
How old is the "act as" prompt?
The pattern dates to December 2022. The GitHub repo f/awesome-chatgpt-prompts was created on 2022-12-05, and its first README already had "Act as a Linux Terminal." Its current prompts file has 946 matches for "act as," including "I want you to act as a linux terminal." The pattern spread through one widely copied list, without research behind it, which fits the EMNLP finding.
Anthropic's current guidance also cuts against longer stacked prompts. It notes recent models "are also more responsive to the system prompt than previous models," and suggests replacing "CRITICAL: You MUST use this tool when..." with "Use this tool when..." The vendor advice favors shorter, plainer instructions.
Does it work better?
The video offers no test. It names no model, benchmark or before-and-after example. Jones closes with a personal note, "I find that very useful when I don't know a ton about a particular space," and "let me know how that goes." The step-back instruction has research behind it, and asking the model to name required tools can surface things you did not know to ask. The claim that the full prompt "works better" than the 2022 version is untested.
Bottom line
Keep the root-cause instruction, which matches Step-Back Prompting and its measured gains. The expert persona line mainly shapes tone. Drop the request to consult training data and give the model a source or a search tool instead.
Sources
- When "A Helpful Assistant" Is Not Really Helpful, EMNLP 2024: personas do not improve factual accuracy.
- Take a Step Back, arXiv:2310.06117: Step-Back Prompting and its reported gains.
- Claude prompting best practices: role prompting, the medication dosing block and overtriggering guidance.
- Equipping agents with Agent Skills, Anthropic: the Agent Skills definition.
- Agent Skills overview: the open standard.
- f/awesome-chatgpt-prompts: the origin of the "act as" pattern.
- Nate B. Jones, About: the creator's name and background.
Published October 21, 2026. Based on a TikTok by @nate.b.jones. Full video notes (markdown)