<- all tokdocs
2026.10.07@jakevanclief

Claude Code's own glossary backs this video's harness argument, and GitHub's docs show Copilot was filed under the wrong heading

AI
Watch on TikTok

View on TikTok ->

The artifact is a 111.618-second MP4 (container mov,mp4,m4a,3gp,3g2,mj2), 720x1280 at 30 fps, HEVC Main profile video in yuv420p tagged hvc1 at about 71 kbps, with an HE-AACv2 stereo audio track at 44.1 kHz and about 64 kbps, for an overall 141 kbps and a 1,973,706-byte file; yt-dlp recorded the format as bytevc1_720p_141460-1. It was posted at 2026-10-07T13:02:58 UTC by uploader handle jakevanclief, channel nickname "Jake Van Clief," over an audio track credited as "original sound" by Jake Van Clief. The description runs 709 characters across 120 words and ends with a Skool link and the hashtags #shorts #ai #aiagents #claudecode #automation. At capture on 2026-10-08 the post had 1,006 views, 32 likes, 2 comments, 1 repost, and 15 saves, which is a 3.2 percent like-to-view rate on a post less than a day old. I read the full 351-word transcript and 17 of the 56 extracted frames (001, 004, 008, 012, 016, 020, 023, 026, 029, 030, 034, 038, 042, 046, 049, 052, 056), sampled evenly because the chalkboard scenes change slowly. There is no live-action footage. The whole video is a pixel-art classroom: a dark maroon brick wall, a green chalkboard in a brown frame with two chalk sticks on the ledge, and a pixel avatar in a white tank top, olive cargo pants, long dark hair, beard, arm tattoos, and a blue pendant, labeled with a small tag reading "JAKE." Burned-in captions sit in a boxed bar at the bottom with the keyword in orange, for example "me 120 grand," "a good harness," "One good model," and "person checks it." The chalkboard content advances through a three-tab progress bar across the top reading "1 WHEEL," "2 3 PARTS," "3 PROOF," with orange checkmarks appearing as each tab completes. An angled orange banner reading "$120K · DUBAI · 3 WEEKS" opens and closes the video. Panel 1, "1 · The wheel is done," shows a chalk wagon wheel labeled "built on what works" beside a square, triangle, and oval each struck through in red under the label "new wheels." Panel 2, "2 · Three parts," stacks three boxed rows: "INSTRUCTIONS" with "what + how," "THE MODEL" with chalk pills reading "Claude," "ChatGPT," and "Copilot," and "WHAT IT CAN REACH" with a document icon and a gear icon, with "HARNESS (Claude Code)" written underneath. Frames 029 and 030 cut to a screen recording of a GitHub page at github.com/RinDig/Interpretable-Context-Methodology showing main, 3 branches, 0 tags, 21 commits, a file list containing _core, workspaces, .gitignore, CLAUDE.md, LICENSE, and README.md with commit messages including "Initial release: MWP framework with 3 workspaces" and "Rename MWP to ICM in CLAUDE.md," and a rendered README headed "Interpretable Context Methodology (ICM)" with the line "Folder structure as agent architecture," a link labeled https://arxiv.org/abs/2603.16021, and the byline "Created by Jake Van Clief," overlaid with an orange banner reading "A GOOD SET OF FOLDERS." Panel 3, "3 · Dubai," draws a document labeled "the brief" with an arrow to a box reading "CLAUDE" struck through in red under "can't reach the site," and a second arrow to a box reading "CHATGPT" with an orange checkmark. A later frame shows a white news card reading "Gartner Warns 70% of Vendor-Built AI Agent Projects Face Abandonment by 2028" credited on screen to "techstrong.ai, reporting Gartner," stamped with an orange "70% · BY 2028" banner. The closing chalkboard lists a four-item checklist: "1. the outcome you want," "2. how you'd do it," "3. what it can touch," "4. where a person checks."

Anthropic's own glossary states the harness-and-model split the video draws on the chalkboard

The video's structural claim is that a model is rented, a harness supplies the looping and tool calls, and the durable asset is the written instructions. Anthropic's Claude Code documentation says the same thing in its own words. The Claude Code glossary defines an agentic harness as "the tools, context management, and execution environment that turn a language model into a capable coding agent," and states plainly: "Claude Code is the harness; Claude is the model inside it. The harness supplies file access, shell execution, permission gating, memory loading, and the loop that chains actions together." The companion page How Claude Code works defines the agentic loop as three blended phases, "gather context, take action, and verify results," and says "Claude Code is the layer around the model that provides the tools and manages the context the model sees. This surrounding layer is what the term agentic harness refers to."

That is a rare case of a short-form video matching vendor documentation on a technical point. The gap worth noting is scope. The docs describe the harness handling the loop for coding and command-line work. They do not claim the harness solves orchestration for every agent workload, and the video's line that "for most people, the agent problem is solved" is the speaker's generalization, not a documented claim.

Two of the three items in the chalkboard's "THE MODEL" row are products, not models

Panel 2 puts three chalk pills in the row labeled "THE MODEL": Claude, ChatGPT, and Copilot. Only one of those names a model family. GitHub's own documentation describes GitHub Copilot as "an AI assistant that helps you write, understand, and ship software," a product rather than a model. GitHub's supported models reference opens with "GitHub Copilot supports multiple AI models, each with different strengths," and as listed on 2026-10-08 that picker spans Anthropic, OpenAI, Google, xAI, Microsoft, and Moonshot AI models. ChatGPT is likewise an interface; the models underneath it carry GPT version numbers.

This matters because it undercuts the video's own argument while it is being made. Copilot belongs in the harness row the video draws two lines lower, next to Claude Code. Putting a harness in the model slot is the exact category confusion the video says costs companies money. Model lineups also go stale within weeks, so treat any specific version list, including the one above, as a snapshot to re-check against the vendor page rather than a fact to memorize.

The Dubai anecdote blames the model for something the docs treat as a reachability problem

The transcript says Claude "couldn't reach the government website it needed," so the same brief was run in ChatGPT and worked. The chalkboard reinforces this with a red strike through "CLAUDE" labeled "can't reach the site."

Partial correction. Claude Code ships web access as a built-in tool category. The How Claude Code works tool table lists a Web row whose stated purpose is "Search the web, fetch documentation, look up error messages," and the glossary's definition of a tool includes "search the web." A failure to load one specific government site is a reachability outcome, driven by permission rules, a corporate proxy, managed settings, or the site itself refusing the request. It is not evidence that the model lacks web capability.

The video's broader point survives the correction and is arguably strengthened. If the brief was portable enough to move between two different systems and still produce the result, the brief was the asset. That is the claim the anecdote actually supports.

The Gartner figure is real, and it covers vendor-assisted builds rather than all agent projects

The transcript says "Gartner's saying up to 70% of companies will walk away from the agents vendors built for them by 2028," and the on-screen card credits techstrong.ai reporting Gartner. That headline exists. Techstrong.ai published it on 2026-09-30 with the wording "By 2028, up to 70% of enterprises will abandon agentic AI systems built through vendor-assisted models as maintenance expenses soar and internal teams struggle to modify the technology independently," quoting Gartner Senior Director Analyst Mukul Saha. The article names forward-deployed engineering, where a vendor embeds its own engineers inside a customer to build custom software, as the mechanism behind the prediction.

Two clarifications. First, the prediction is scoped to systems built through vendor-assisted engagements, so it is not a forecast that 70 percent of all agent projects fail. The video's phrasing, "the agents vendors built for them," stays inside that scope, which is more careful than most repetitions of this stat. Second, this is a different and newer prediction than the one it is routinely confused with. Gartner's June 25, 2025 press release predicts that over 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls, and it introduces the term "agent washing" for rebranded chatbots and RPA. The 40 percent by 2027 figure and the 70 percent by 2028 figure are separate claims with different scopes. The techstrong.ai article is trade reporting on a Gartner report rather than the Gartner press release itself.

The GitHub repo on screen is real, and the paper behind it still calls the method MWP

github.com/RinDig/Interpretable-Context-Methodology exists, is MIT licensed, carries the 21 commits visible in frame 029, and credits Jake Van Clief. Its README opens with "Folder structure as agent architecture" and describes ICM as replacing framework-level orchestration with filesystem structure, where numbered folders represent stages and markdown files carry the prompts.

The linked paper is arXiv:2603.16021, titled "Interpretable Context Methodology: Folder Structure as Agentic Architecture," filed under cs.AI and cs.HC, submitted as v1 on 2026-03-17 and v2 on 2026-03-18. Two details the video leaves out. The paper has a second author, David McDermott, who is never mentioned. And the abstract still presents the system as the Model Workspace Protocol (MWP), not ICM, which lines up with the repo's own commit history visible on screen: "Initial release: MWP framework with 3 workspaces" followed by "Rename MWP to ICM in CLAUDE.md." The rename came after publication. An arXiv posting is a preprint, so being on arXiv establishes that the method was written up and timestamped, not that it was peer reviewed.

The $120,000, the Dubai location, and the three-week duration have no public record

The entire framing of the video rests on a client engagement. The opening line is "A company paid me $120,000 to teach their people that agents are just a naming convention," and the orange banner adds "DUBAI" and "3 WEEKS." The three-week figure appears only on screen and is never spoken.

No client is named, no contract is shown, and no third-party source corroborates the fee, the location, or the duration. Treat all three as self-reported and unverified. The creator's public footprint supports the general shape of the story without confirming its specifics: the GitHub repo, the arXiv preprint, and the Skool community at skool.com/cliefnotes linked verbatim in the video description are all real and consistent with someone who runs training engagements. Biographical details circulating about him, including military service, a postgraduate degree, and named Fortune 500 clients, come from promotional community-review sites rather than primary records, so they carry the same unverified status as the fee.

One more unsourced claim deserves flagging. The transcript asserts "everybody's out here building their own AI agent from scratch. Like, everybody." No survey, repo count, or vendor figure is offered. The Gartner material cited later in the same video actually describes the opposite failure mode, which is enterprises buying agents from vendors and then abandoning them.

Key Takeaways

  • Verified: The harness-and-model distinction the video builds its argument on matches Anthropic's documentation word for word. The Claude Code glossary states "Claude Code is the harness; Claude is the model inside it," and defines the agentic loop as gather context, take action, verify results.
  • Verified: The GitHub repo, the arXiv preprint, and the Gartner headline shown on screen all exist as displayed, including the 21 commits and the MIT license.
  • Correction: Copilot is not a model. GitHub documents it as an AI assistant whose model picker spans Anthropic, OpenAI, Google, xAI, Microsoft, and Moonshot AI models. It belongs in the harness row the video draws two lines below, alongside Claude Code. ChatGPT is also a product rather than a model family.
  • Partial correction: "Claude couldn't reach the government website" is presented as a model limitation. Claude Code's documented tool list includes a Web category for searching the web and fetching documentation. A single site failing to load is a permissions, proxy, or site-blocking outcome, not a missing capability.
  • Clarification: The Gartner 70 percent figure is scoped to agentic AI systems built through vendor-assisted engagements such as forward-deployed engineering. It is a separate prediction from Gartner's June 2025 forecast that over 40 percent of agentic AI projects will be canceled by the end of 2027.
  • Clarification: The arXiv abstract still names the method Model Workspace Protocol while the title and repo say Interpretable Context Methodology, and the paper lists a co-author, David McDermott, who the video does not mention.
  • Context: The $120,000 fee, the Dubai setting, and the three-week duration are self-reported with no public record. The three-week figure appears only in the on-screen banner and is never spoken.
  • Context: "Everybody's out here building their own AI agent from scratch" is unsourced, and the Gartner research cited later in the video describes enterprises abandoning vendor-built agents rather than over-building their own.

Resources

Published October 7, 2026. Writeup generated from a favorited TikTok.