<- all tokdocs
2026.10.21@nate.b.jones

Nate B. Jones frames this as an upgraded expert prompt, but the measured gains belong to its root-cause instruction, and its "go back and look at your training data" step describes something no model does

AI
Watch on TikTok

View on TikTok ->

The artifact is a 115.45 s vertical clip, 1080x1920 at 0.56 aspect, SDR, in an MP4 container that ffprobe reports as mov,mp4,m4a,3gp,3g2,mj2, 23,477,693 bytes, overall bitrate 1,626,892 bps; the video stream is HEVC Main profile level 150, yuv420p, nominal 30/1 frame rate with an average of 103860/3463 (about 29.99 fps), 3,462 frames, 1,524,266 bps, and the audio stream is HE-AAC, 44.1 kHz, 2 channels, 96,117 bps, 2,489 frames, with container duration 115.448005 s and the extracted audio.mp3 measuring 115.513469 s; yt-dlp recorded format id bytevc1_1080p_1626878-1, uploader nate.b.jones with uploader id 7061010091441964079, channel name "Nate", and audio credited to "original sound" by artist "Nate"; the post timestamp is epoch 1792626649, which is 2026-10-21 23:50:49 UTC, matching the recorded upload_date of 20261021; the description is 79 characters long and reads "Let me know how it goes. #aihack #ai #aiprompts #artificialintelligence #aitips", and at capture time, the 2026-10-22 12:05:58 local mtime of metadata.json, the counters stood at 75,800 views, 5,544 likes, 136 comments, 970 reposts and 4,865 saves, a save count 35.8 times the comment count and equal to 6.4 percent of views, which is the signature of a copy-this-prompt post rather than a debate post; I read 12 of the 58 extracted frames, numbers 001, 005, 008, 012, 018, 025, 032, 038, 044, 050, 055 and 058, which at the 2 s extraction interval sit at roughly 0, 8, 14, 22, 34, 48, 62, 74, 86, 98, 108 and 114 seconds, and every one of them shows the same single unbroken shot: a bearded man in dark-rimmed glasses and a dark crew-neck shirt, seated against a pale olive chair back with a grey woodgrain wall behind him, gesturing at the lens, with a blue rounded title card pinned to the top third reading "AI Hack: The Expert Prompt, Upgraded For 2026" in white, present and unchanged in all 12 frames, and a burned-in word-level caption strip near the bottom that reads "SO MUCH OF" in 001, "IT'S THAT AI" in 005, "A SECRET" in 008, "DOING WITH AI" in 012, "KNOW A LOT" in 018, "LOOK AT THE" in 032, "OR SKILLS INVOLVED" in 038, "AND THEN COME" in 044, "CAUSE I'M NOT" in 050, "BUT THIS IS" in 055 and "ABOUT A PARTICULAR" in 058, with frame 025 showing no caption at all while he holds a finger to his lips during the pause before he delivers the prompt; no screen recording, no chat window, no terminal, no editor, no browser, no URL, no repo name, no version number and no chart appears in any frame I read, and the prompt he recites is never shown as text on screen, so a viewer who wants it has to transcribe it from speech; the spoken argument across 25 transcript segments and 400 words is that people ask AI for work outside their own expertise, that the bottleneck is the asker rather than the model, and that a four-part instruction (act like an expert, go back to root causes and restate the problem, propose the simplest solution, name the tools and skills required) repairs the gap, which he describes at 105.28 s as an update to "the old, you're an expert prompt"; the transcript-end check passes cleanly, because the final segment runs 112.38 s to 115.20 s against a real audio duration of 115.513469 s, so every segment lies wholly inside the audio and nothing here is a Whisper overshoot, and the usual product-name mangling risk does not apply because the speaker names no product, company or person anywhere in the 400 words, referring only to "AI" and "the AI" throughout.

The "act like an expert in this domain" opener is the one sentence in the script with no measured effect

At 48.84 s the video says "You need to act like an expert in this domain." That line is positioned as the hinge of the whole technique, and at 105.28 s it is explicitly tied to lineage: "In a sense, we're updating the old, you're an expert prompt, but this is now for 2026 and it works better."

The persona half of that sentence is the part the literature tests directly, and the result is unflattering. Mingqian Zheng, Jiaxin Pei, Lajanugen Logeswaran, Moontae Lee and David Jurgens published "When 'A Helpful Assistant' Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models" in Findings of the Association for Computational Linguistics: EMNLP 2024. Their abstract states that they "curate a list of 162 roles covering 6 types of interpersonal relationships and 8 domains of expertise" and that "through extensive analysis of 4 popular families of LLMs and 2,410 factual questions, we demonstrate 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." They go further on the search question, reporting that "while aggregating results from the best persona for each question significantly improves prediction accuracy, automatically identifying the best persona is challenging, with predictions often performing no better than random selection."

Anthropic's own prompting documentation makes a narrower claim than the video does. The "Give Claude a role" section of the Claude prompting best practices page says "Setting a role in the system prompt focuses Claude's behavior and tone for your use case. Even a single sentence makes a difference." The claimed benefit is behavior and tone. The page does not claim that a role improves factual accuracy, which is the thing the video is implicitly promising when it says the prompt will tell you "what a proper expert would do."

So the correction is narrow and worth stating precisely. The script's opening line is the one part of it that a published, dated study found had no reliable effect on objective accuracy. The three instructions that follow it are doing the work.

"Go back to root causes, fully identify the problem" is the part with measured evidence behind it

At 51.64 s the video says "And to me, that means that you should go back to root causes, fully identify the problem, even if I haven't expressed it correctly," and it restates the same idea at 62.54 s as "Look at the root causes in the problem that I'm talking about."

That instruction has a published analogue with numbers attached. Huaixiu Steven Zheng, Swaroop Mishra, Xinyun Chen, Heng-Tze Cheng, Ed H. Chi, Quoc V Le and Denny Zhou submitted "Take a Step Back: Evoking Reasoning via Abstraction in Large Language Models" to arXiv on 9 October 2023 as arXiv:2310.06117, and it was accepted at ICLR 2024. The paper 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." The reported gains on PaLM-2L are 7 percent on MMLU Physics, 11 percent on MMLU Chemistry, 27 percent on TimeQA and 7 percent on MuSiQue, with experiments also run on GPT-4 and Llama2-70B.

The mechanism described in the paper and the mechanism described at 51.64 s are the same move. Both ask the model to abandon the user's framing of the particular question and reconstruct the general problem underneath it. The video's phrase "even if I haven't expressed it correctly" is a plain-language statement of the abstraction step, and at 94.32 s the video reaches the same conclusion from the other direction: "you're basically telling the AI, don't get locked into what I said because I'm not the expert."

This is the strongest part of the video and it is the part the video does not label as the important part. The title card calls the technique "The Expert Prompt." The evidence attaches to the step-back instruction underneath the persona line.

No model can "go back and look" at its training data, and Anthropic's own docs instruct the opposite

At 99.90 s the video says "Please, like go back, you have training data, go back and look and see what a proper expert would do with this."

This describes an operation that does not happen. A deployed language model has fixed weights at inference time and no index over the corpus it was trained on. There is no lookup to perform, no corpus to consult and no way for the model to distinguish a fact it absorbed from one it is reconstructing. Asking it to "go back and look" shifts which tokens it finds probable. It does not initiate retrieval. Retrieval requires an attached tool, and the video attaches none.

Anthropic's prompting documentation encodes the opposite instruction in the one place where the stakes are highest. Its medication dosing guidance explains that the recommended system prompt block "has Claude give a child's dose of an over-the-counter medicine from the product's label, not from memory," and the block itself reads in part: "If web search is available, Claude uses it to find the product's label and confirms it against an official copy of the same label, such as DailyMed. If web search is not available or Claude can't find the label, 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."

The practical consequence matters for the use cases the video names. At 35.96 s it suggests the technique for someone "having trouble with real estate listings and you're looking for a house and you don't know a lot about that." Real estate figures are local, dated and frequently revised. A prompt that instructs the model to consult its training data will return a confident answer assembled from whatever the weights encode, with no way for a non-expert reader to tell a current number from a stale one. The fix is a search tool or a pasted source document, and the prompt in this video asks for neither.

"Tools or skills" now names a specific artifact, and a chat model cannot install one

At 71.34 s the video says "And make sure that if there are tools or skills involved that I need to use or let you have to get this done, that you find the simplest, clearest possible configuration of skills and tools and say, this is what I need to solve this problem for you." Frame 038, at roughly 74 s, carries the burned-in caption "OR SKILLS INVOLVED" over that line.

As of the post date, "skills" carries a defined technical meaning in this exact context. Anthropic published "Equipping agents with Agent Skills" on its engineering blog on 16 October 2025, defining Agent Skills as "organized folders of instructions, scripts, and resources that agents can discover and load dynamically to perform better at specific tasks," and specifying that at minimum a skill is a directory containing a SKILL.md file that "must start with YAML frontmatter that contains some required metadata: name and description." The format was subsequently released as an open standard, and the agentskills.io specification site describes it as "a lightweight, open format for extending AI agent capabilities with specialized knowledge and workflows" that loads through "progressive disclosure" in three stages, discovery, activation and execution, noting that "the Agent Skills format was originally developed by Anthropic, released as an open standard, and has been adopted by a growing number of agent products."

The video uses "skills" in the older colloquial sense of capabilities a person has, not in the Agent Skills sense. That is fair usage, but it creates a gap between the ask and the result. The prompt instructs the model to "find the simplest, clearest possible configuration of skills and tools." A plain chat session has no mechanism to install, enable or verify any of them. The output is a shopping list the reader then has to act on by hand. The video does not say this, and no frame I read shows a tool being added or a skill being loaded.

The "old, you're an expert prompt" is datable to the first week of December 2022

At 105.28 s the video says "In a sense, we're updating the old, you're an expert prompt, but this is now for 2026 and it works better." The "old" part of that is checkable and it holds up.

The canonical collection of that prompt pattern is the GitHub repository f/awesome-chatgpt-prompts, now operating as prompts.chat. The GitHub API reports its created_at as 2022-12-05T13:54:13Z. The README as it stood at commit 17b4aa75, the version in place within a day of repository creation, already contained the heading "## Act as a Linux Terminal." The current prompts.csv in that repository is 7,774,594 bytes and contains 946 case-insensitive matches for the string "act as," with entries that open exactly the way the video describes, for example "I want you to act as a linux terminal" and "I want you to act as an English translator, spelling corrector and improver." The repository currently shows 173,966 stars.

So the pattern the video calls old is just under four years old at the 2026-10-21 post date, and it was popularized through a single widely copied artifact rather than emerging from research. That is useful context for the "upgraded" framing: the thing being upgraded was never a validated technique in the first place, which is consistent with the EMNLP 2024 finding that it does not move accuracy.

One further wrinkle sits against the "2026" framing. Anthropic's current prompting best practices page warns that recent models "are also more responsive to the system prompt than previous models," and that prompts written to force tool or skill usage "might now overtrigger," with the recommended fix being to soften emphasis: "Where you might have said 'CRITICAL: You MUST use this tool when...,' you can use more normal prompting such as 'Use this tool when...' instead." The documented 2026-era direction of travel is toward shorter and plainer instructions, not toward longer stacked ones. The video's prompt is a single long sentence stack, which is the pattern the vendor documentation is now steering away from.

Whether it "works better" is unverified, and the video runs no test

At 89.18 s the video says "And this works for a wide range of stuff," and at 105.28 s it says "it works better." At 112.38 s, the final segment, it retreats to a personal report: "But I find that very useful when I don't know a ton about a particular space."

No evidence supports the comparative. The video names no model, no benchmark, no before-and-after pair and no sample size. Across all 12 frames I read, spanning 0 s to 114 s, there is no chat interface, no model output, no side-by-side comparison and no screen recording of any kind. The entire runtime is a single talking-head shot. The closing line at 111.18 s, "So let me know how that goes," hands the evaluation to the audience, which is also where the title of the post comes from.

That does not make the prompt bad. The step-back instruction inside it has published support, and the request to name required tools and skills is a sensible way to surface unknown unknowns. It does mean the specific claim "it works better" than the 2022 persona prompt is an untested assertion in this artifact, and the 4,865 saves recorded at capture time are people banking an untested instruction rather than a demonstrated result.

Key Takeaways

  • Correction: The video says at 99.90 s "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 over its training corpus, so no lookup occurs. Anthropic's own prompting documentation instructs the reverse for high-stakes answers, stating that Claude "doesn't give a number from memory" and should confirm against a retrieved source such as DailyMed when web search is available.
  • Correction: The video presents "act like an expert in this domain" at 48.84 s as the core of the technique. Zheng, Pei, Logeswaran, Lee and Jurgens tested 162 roles across 4 LLM families and 2,410 factual questions in Findings of EMNLP 2024 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."
  • Verified: The root-cause instruction at 51.64 s matches a published technique with measured gains. Step-Back Prompting (arXiv:2310.06117, submitted 9 October 2023, ICLR 2024) reported improvements on PaLM-2L of 7 percent on MMLU Physics, 11 percent on MMLU Chemistry, 27 percent on TimeQA and 7 percent on MuSiQue.
  • Verified: The "old, you're an expert prompt" the video references at 105.28 s is datable. The f/awesome-chatgpt-prompts repository was created 2022-12-05T13:54:13Z, its earliest README already contained "Act as a Linux Terminal," and its current prompts.csv carries 946 matches for "act as."
  • Clarification: "Skills" at 71.34 s is used colloquially, but since Anthropic's 16 October 2025 engineering post the word names a specific artifact: a folder containing a SKILL.md with name and description in YAML frontmatter. A plain chat session cannot install or enable one, so the prompt's output on that point is a list the reader must act on manually.
  • Partial correction: The video positions longer, stacked instructions as the 2026 upgrade. Anthropic's current prompting best practices page notes that recent models are "more responsive to the system prompt than previous models," that forceful phrasing can cause overtriggering, and recommends softening "CRITICAL: You MUST use this tool when..." to "Use this tool when..."
  • Unverified: "This works for a wide range of stuff" (89.18 s) and "it works better" (105.28 s) are supported by no test in the artifact. No model is named, no benchmark is cited, and none of the 12 frames I read contains a chat window, model output or comparison.
  • Context: The artifact is a single unbroken talking-head shot, 115.45 s, HEVC 1080x1920, with a fixed title card reading "AI Hack: The Expert Prompt, Upgraded For 2026" and burned-in word-level captions. The recited prompt is never rendered as on-screen text.
  • Context: At capture time the post showed 75,800 views, 5,544 likes, 136 comments, 970 reposts and 4,865 saves. Saves outnumber comments 35.8 to 1, which fits a prompt people bank rather than argue about.
  • Context: The transcript is clean. The last of 25 segments ends at 115.20 s against an audio duration of 115.513469 s, so no segment extends past the audio, and the speaker names no product, company or person across the 400 words, removing the usual Whisper name-mangling risk.

Resources

Published October 21, 2026. Writeup generated from a favorited TikTok.