Claire Vo uses a 27-minute How I AI solo episode to show what a Claude Skill actually contains and how she creates one. Published around the feature’s October 2025 launch, the walkthrough follows the rough edges as well as the successful path. She tries Claude’s web-based creator, switches to Cursor for local file creation, tests the result in Claude Code, and packages another skill for the web interface.

Package a recurring task · 01:40–04:57

  • Vo defines a skill as reusable instructions and context for a particular job. Her examples include analyzing data, creating a document, following a workflow, and running a script.
  • She contrasts that task-specific package with the broader context she keeps in Projects and custom GPTs. The attraction is discovering the right procedure within a conversation, without manually pasting a saved prompt each time.
  • She favors instructions written in natural language over assembling every task as a rigid visual automation. That preference still leaves room for explicit steps, examples, and executable code.
  • A skill therefore gives her existing prompt collection a structure. The unit of reuse is the procedure, with a defined input and expected output.

The artifact is a folder · 05:00–10:59

  • A skill contains a main SKILL.md file and, when useful, supporting files. Vo emphasizes this concrete packaging because the feature’s abstract explanations initially left her unsure what she was supposed to create.
  • The main file starts with YAML metadata, structured fields for the skill’s name and description. Its body contains Markdown instructions. Relative links point to examples, templates, further guidance, or scripts alongside it.
  • Those links let the agent bring in supporting context as needed. The folder can hold more than a single large prompt while keeping the core procedure legible.
  • Python scripts can supply a consistent operation, such as cleaning data or validating files. Vo contrasts reusing a known script with asking a model to generate a fresh implementation every time.
  • Local files can be used from Claude Code; the web workflow shown later requires a ZIP upload. Anthropic’s explanation of Agent Skills confirms the SKILL.md filename and folder structure.

Learn from the web creator’s output · 11:02–14:54

  • Vo asks Claude’s web app to help create skills, then requests one for generating a product requirements document, a brief describing a feature.
  • The generated instructions include when to use the skill, decision points, document templates, writing guidance, examples, and questions to ask the user. She finds the clarifying questions especially useful as a prompting pattern.
  • She also notices keywords, but explicitly treats their role in invoking the skill as a guess. The episode does not establish a required keyword mechanism.
  • Her recommendation is to inspect one generated example even if this is not the authoring route you ultimately choose. It provides a concrete model of a detailed procedure.

Keep authoring close to the files · 14:56–17:39

  • The web creator produces 12 files where Vo believes five would have sufficed. It adds material such as a license and a quick-reference guide, then fails when she tries to download the files.
  • She switches to an empty local folder opened in Cursor. Her prompt asks for a skill that creates skills and includes the official documentation as the specification.
  • Cursor creates the main instructions, examples, templates, and a validation script in a folder. Vo estimates roughly three minutes for this route versus roughly ten for the earlier web attempt.
  • Those timings describe her experience in the episode. The practical difference is that the files already exist locally, avoiding the failed download handoff.

Validate the package, then test the task · 17:42–20:51

  • The generated Python validator checks YAML, formatting, file existence, and some content. It gives Vo a repeatable check of the package’s structure.
  • She launches Claude Code in the local directory and asks it to use the creator skill to make another skill: turn technical changelog entries into a customer-facing newsletter.
  • This is a recurring ChatPRD job with a specific input, output, and format. Claude Code finds and reads the creator instructions, makes the newsletter skill’s directory, writes the file, and runs validation.
  • The new skill needs only one instruction file. Vo points out that linked examples and extra folders are optional, not a requirement to make every skill elaborate.

Test discovery and revise the instructions · 20:54–22:14

  • Vo supplies the latest changelog without explicitly requesting a newsletter or naming the skill. In this conversation, Claude infers the task and generates an October 2025 product update.
  • She immediately identifies a style issue: the output uses too many emoji. Her next improvement would be to revise the reusable instructions until the newsletter matches her desired format.
  • The example separates successful discovery from satisfactory output. Finding the skill does not remove the need to inspect the result and improve its guidance.

Upload, fix the name, and preserve the correction · 22:16–26:31

  • Vo requests a second skill for turning customer-demo notes into personalized follow-up emails for trial prospects. Claude Code creates it and runs the validator again.
  • She then compresses the package and uploads it under the web app’s capabilities settings. The first attempt fails because of the skill name; she changes it to the lowercase, hyphenated form accepted in the demo.
  • She also updates the creator skill so future packages inherit that correction. The retry succeeds and the uploaded skill becomes available in chat.
  • The episode ends with the follow-up skill ready to receive demo notes. It shows creation and upload, rather than a completed prospect-email example or an API integration.

Caveat: this captures a launch-era October 2025 workflow; today’s installation paths, naming rules, and interface details may differ.