<- all tokdocs

Sharing a Claude Skill With Your Team: The SBAR Report Example

By , founder of Mavric and Apso

View on TikTok ->

A TikTok from @jenna_gardner_ai describes building an "SBAR Claude skill" that scores a team's paid media reports from one to five before they reach a senior reviewer. A Claude Skill is a sound way to package a shared rubric like this. Two parts of the story need caveats: how the Skill gets shared with a team depends on which Claude surface they use, and the numeric score is the least reliable output.

The clip is a talking-head video with no screen recording, and only the audio names the tool.

What is an SBAR Claude Skill?

She describes a colleague whose team "does SBAR reporting," meaning their paid media performance reports follow a set structure. Instead of a senior person fixing every report, she and the colleague built a Skill that checks drafts against that structure:

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

SBAR stands for Situation, Background, Assessment, Recommendation. The Institute for Healthcare Improvement 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." Its documented home is clinical handoff. Using it for paid media reviews is a reasonable transplant, though I found no source that establishes it as standard practice in agency reporting.

How Claude Skills work

Per Anthropic's Agent Skills documentation, a Skill is a directory with a required SKILL.md file. Its YAML front matter has two required fields, name and description. Claude loads a Skill in three levels:

  1. Metadata is always in the system prompt, at roughly 100 tokens per Skill.
  2. The SKILL.md body loads only when the description matches the request. The docs budget it at under 5k tokens.
  3. Extra resources and scripts cost nothing until Claude reads them.

That design is why a shared rubric fits a Skill better than a pasted prompt. Nobody on the team 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 case: "Claude is powerful, but real work requires procedural knowledge and organizational context."

Can you name a Skill "Claude"?

No. The docs say a Skill's name field "cannot contain reserved words: 'anthropic', 'claude'". An "SBAR Claude skill" would need a name like sbar-review.

How to share a Claude Skill with your team

The payoff line in the clip 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

How hard that is depends on the surface, which she never names. The Agent Skills documentation spells out 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 live on the filesystem, in ~/.claude/skills/ for personal use or .claude/skills/ for a project, and can be shared through Claude Code Plugins.

The docs also say "custom Skills do not sync across surfaces." If this team works in claude.ai, "used by her team" means every person imported the same folder by hand. There is no version control and no admin push, so each update has to be redistributed manually.

Are LLM scores from 1 to 5 reliable?

The research says they drift between runs. 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," calling the ratings "almost arbitrary in the worst case." Running the same draft through the same Skill twice can produce two different numbers.

Her choice of scale is close to the best option. Grading Scale Impact on LLM-as-a-Judge, dated 2026-01-08, found that a 0-5 scale produced the strongest agreement between human and model raters among the scales tested. A 1-to-5 scale is one step away.

The more useful output is the one she mentions almost in passing: "where are the parts specifically that it needs to be improved and how to strengthen it." Specific, located feedback holds up across runs in a way a single integer does not.

Why fixing work upstream works

The thesis of the clip 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, as listed by the University of Cambridge Institute for Manufacturing, reads: "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 is inspection on a mass basis. Giving the rubric to the people who write the reports builds quality in.

She also says running the rubric teaches the team "what good looks like." A Skill that returns explained feedback trains the writer, while one that returns only a number does not. One risk the clip does not cover is a team that learns to write for the rubric instead of the client, which raises scores without improving reports.

Bottom line

A Claude Skill is the right container for a shared review rubric like SBAR. On claude.ai, each teammate has to upload it separately, and updates do not sync. Treat the 1-to-5 score as rough and rely on the specific feedback.

Sources

Published September 5, 2026. Based on a TikTok by @jenna_gardner_ai. Full video notes (markdown)