<- all tokdocs
2026.09.05@jenna_gardner_ai

The "fix it upstream" clip shows no automation canvas, and the Claude Skill it describes cannot be distributed to a team on claude.ai

AI
Watch on TikTok

View on TikTok ->

A 103-second clip (1:43), 1080x1920 vertical, uploaded 2026-09-05 by @jenna_gardner_ai on the channel jenna_gardner_ai. At capture the counts were 3,206 views, 152 likes, 19 comments, 12 reposts, and 33 saves. The audio is listed as "original sound" by jenna_gardner_ai, so there is no licensed music bed. I read all 52 frames in tokdoc_frames/ and the 334-word Whisper transcript. The frames contain no software at all. Every frame is a front-facing selfie shot in a home kitchen and dining area: pendant lights overhead, a counter with glass jars and a small pink animal figurine, a framed picture on the left wall, dark cabinet doors behind. She wears a black crewneck whose white chest text is mirrored by the front camera, reading "CARDI" over "SURF C" in reverse. A static black-box hook caption, "How to save yourself from fixing other people's work," sits at the top of frames 1 through 8, roughly the first 15 seconds, then disappears for the rest of the clip. Karaoke-style burned-in captions run the full length. There is no workflow canvas, no node, no logo, no prompt text, no score readout, no pricing figure, and no watermark beyond the captions.

The platform question has a negative answer, and it matters

The brief for this analysis assumed an automation platform on screen. There is none. Across all 52 frames there is no n8n canvas, no Make scenario, no Zapier zap, no screen recording of any kind. The only "platform" named is in the audio, and only by ear.

Whisper renders it as a misspelling:

and so we created a clod skill we're like all right we're gonna make an SBAR clod skill

"Clod" is the ASR rendering of "Claude." That reading is supported by the noun that follows. "Skill" is the exact product term Anthropic uses for a packaged instruction bundle, documented in Agent Skills. No other vendor ships a first-class artifact called a Skill that a non-engineer builds in a one-on-one session.

The word "Claude" never appears in writing anywhere in the video. The burned-in caption at frame 23 reads "WE'RE GONNA MAKE AN S" and cuts off, and the caption at frame 35 reads "SO THAT SKILL IS NOW USED BY HER TEAM." Neither spells the vendor. So the platform identification rests entirely on audio inference, and I am naming it as inference rather than visual evidence.

The description's word "workflow" refers to a human handoff chain, one team drafting reports and a senior person cleaning them up. It does not refer to an automation graph.

SBAR is real, and it comes from clinical handoffs rather than paid media

She is vague on the acronym in the audio:

she said team does SBAR reporting which means that once it's time to report on the paid media performance metrics that they need to have a particular structure on how they report on it as far as like structured communication blah blah blah I don't remember everything

SBAR stands for Situation, Background, Assessment, Recommendation, per the Institute for Healthcare Improvement's SBAR tool page, which describes it as "an easy-to-remember, concrete communication mechanism for framing any conversation, especially critical ones, requiring a clinician's immediate attention and action." Her characterisation of it as structured communication designed to make information land cleanly is correct.

Two caveats. First, SBAR's documented home is clinical handoff, not marketing reporting. Applying it to paid media performance reviews is a reasonable transplant, and I found no primary source establishing it as a standard practice in agency reporting. Second, the frequently repeated origin story (US Navy nuclear submarines, then Kaiser Permanente in 2002) could not be confirmed against a source I was able to fetch. Both the AHRQ TeamSTEPPS SBAR page and the NCBI Bookshelf table that carries the Kaiser attribution returned 403 and CAPTCHA responses to automated retrieval. I am leaving the origin unverified rather than repeating it on secondary authority.

Note also a transcript artifact worth knowing about: Whisper writes the acronym as "SBAR" in the text file, while the burned-in on-screen caption at frame 7 renders the same moment as "DOES S BAR R". Two different automatic captioners, two different spellings of one word.

A Skill is the correct shape for this problem, and the docs explain why

The mechanism she describes maps cleanly onto how Agent Skills actually work. Per Anthropic's Agent Skills documentation, a Skill is a directory containing a required SKILL.md file with YAML frontmatter carrying two required fields, name and description. Loading happens in three stages. Level 1 metadata is always in the system prompt at roughly 100 tokens per Skill. Level 2, the SKILL.md body, loads only when the description matches the request, and the docs budget it at under 5k tokens. Level 3 resources and scripts cost nothing until read.

That architecture is exactly why a shared rubric belongs in a Skill instead of a pasted prompt. The team never has to remember to paste the SBAR criteria. Anthropic's engineering post Equipping agents for the real world with Agent Skills, published 2025-10-16, makes the same argument: "Claude is powerful, but real work requires procedural knowledge and organizational context."

One small and checkable detail cuts against the clip's phrasing. The docs state that a Skill's name field "cannot contain reserved words: 'anthropic', 'claude'". So the artifact she calls an "SBAR Claude skill" cannot literally be named that. It would be sbar-review or similar.

The distribution claim is the weakest link in the story

The core payoff line is a deployment claim:

so that skill is now used by her team and so they all before they send anything to her they're going to run that SBAR get their score and then be able to strengthen

Whether that is easy depends entirely on which Claude surface the team uses, and she never says. The Agent Skills documentation is blunt about the differences:

  • On claude.ai, custom Skills are "individual user only. Each team member must upload separately." The docs add that "claude.ai does not support centralized admin management or org-wide distribution of custom Skills."
  • On the Claude API, custom Skills are workspace-wide and all workspace members can access them.
  • In Claude Code, Skills are filesystem-based, living in ~/.claude/skills/ for personal use or .claude/skills/ for a project, and can be shared through Claude Code Plugins.

The docs also state that "custom Skills do not sync across surfaces." If this team lives in claude.ai, which is the likely home for a marketing team doing one-on-one AI sessions, then "the skill is now used by her team" means someone zipped a folder and every person imported it by hand, with no version control and no admin push. That is a real operational cost the clip skips. The upstream fix she is celebrating has a manual distribution problem sitting directly underneath it.

The 1-to-5 score is the part most likely to disappoint

The specific promise is a numeric quality grade:

it's going to score what quality it is from one to five and where are the parts specifically that it needs to be improved and how to strengthen it

Model-generated numeric ratings are less stable than they look. Rating Roulette: Self-Inconsistency in LLM-As-A-Judge Frameworks by Rajarshi Haldar and Julia Hockenmaier, submitted 2025-10-31 and accepted at EMNLP 2025, finds that LLM judges "have low intra-rater reliability in their assigned scores across different runs," describing the ratings as "almost arbitrary in the worst case." Re-running the same draft through the same Skill can produce a different number.

Her chosen scale is close to the best available choice. Grading Scale Impact on LLM-as-a-Judge: Human-LLM Alignment Is Highest on 0-5 Grading Scale, dated 2026-01-08, finds the 0-5 scale produces the strongest agreement between human and model raters among the scales tested. A 1-to-5 scale is one notch away from that.

The practical read: the number is the least trustworthy output, and the clause she almost throws away, "where are the parts specifically that it needs to be improved and how to strengthen it," is the part that carries the value. Specific located feedback survives run-to-run variance in a way a single integer does not.

No model is named anywhere in the audio or on screen, so there is no out-of-date model claim to flag. For reference, Anthropic's models overview as fetched on 2026-09-27 lists the current lineup as Claude Fable 5.1 ($10 per input MTok, $50 per output MTok), Claude Opus 5.5 ($4 / $20), Claude Sonnet 5 ($2 / $10), and Claude Haiku 4.5 ($1 / $5), with Claude Opus 5 now listed under legacy models still available.

The upstream argument is sound and about eighty years old

Strip the AI framing and the thesis is a quality-management classic:

whoever is kind of mucking up the downstream and that's where she was at we got to fix it upstream

Deming's third point reads, in the University of Cambridge Institute for Manufacturing's listing of the 14 points: "Cease dependence on inspection to achieve quality. Eliminate the need for inspection on a mass basis by building quality into the product in the first place." A senior person rewriting every report before it reaches the client is inspection on a mass basis. Handing the rubric to the drafters is building quality in.

Where the clip earns its keep is the second-order effect she names, that running the rubric teaches the team "what good looks like." That is the difference between a gate and a teaching tool. A Skill that returns located, explained feedback trains the sender. A Skill that returns only a number trains nobody.

The argument does have a failure mode she does not mention. If the drafting team learns to write for the rubric rather than for the client, the score goes up and the reports get worse. Nothing in the clip describes a check on that.

Key Takeaways

  • No automation platform appears in any of the 52 frames. The clip is a talking-head selfie with a static hook caption on frames 1 through 8 and karaoke captions throughout. Identifying the tool as Claude rests on audio inference from the word "skill," not on visual evidence.
  • SBAR means Situation, Background, Assessment, Recommendation, confirmed against the IHI tool page. Its documented home is clinical handoff communication.
  • The Agent Skills mechanism she describes is real and documented: a SKILL.md with required name and description frontmatter, three-level progressive disclosure, roughly 100 tokens per Skill for always-loaded metadata and under 5k for the triggered body.
  • A Skill's name cannot contain the reserved words "claude" or "anthropic," so an "SBAR Claude skill" cannot carry that literal name.
  • The "now used by her team" claim is surface-dependent. On claude.ai, custom Skills are per-user with no org-wide distribution and no admin management, and Skills do not sync across surfaces. On the Claude API they are workspace-wide; in Claude Code they are filesystem-based and shareable through plugins.
  • The 1-to-5 score is the shakiest output. EMNLP 2025 work (arXiv 2510.27106, 2025-10-31) found LLM judges have low intra-rater reliability across runs. A January 2026 paper (arXiv 2601.03444) found 0-5 gives the best human alignment among scales tested.
  • The upstream thesis matches Deming's third point, verbatim as listed by the Cambridge IfM: cease dependence on inspection, build quality in.
  • Transcript artifacts to be aware of: Whisper writes "clod skill" for "Claude Skill," and emits two spurious "you" lines running to 1:53.72 in transcript.srt, past the clip's actual 1:43 runtime.
  • Unverified: the commonly cited SBAR origin story (US Navy submarines, Kaiser Permanente 2002), because the AHRQ TeamSTEPPS page returned 403 and the NCBI Bookshelf attribution table returned a CAPTCHA to automated fetch; whether SBAR is an established paid-media reporting convention, for which I found no primary source; which Claude surface the team uses; the existence, contents, and adoption of the actual SBAR Skill, all of which are private to her client; and the identity of the colleague and the agency, which the clip deliberately withholds.

Resources

Published September 5, 2026. Writeup generated from a favorited TikTok.