How I AIClaude Skills explained: How to create reusable AI workflows
Claire Vo on build reusable Claude workflows using Skills, Cursor, and Claude Code.
In this episode of How I AI, featuring Claire Vo, Claude Skills explained: How to create reusable AI workflows explores build reusable Claude workflows using Skills, Cursor, and Claude Code Claire Vo explains Anthropic’s new Claude Skills feature as a lightweight way to package reusable, task-specific instructions (and optional files/scripts) that Claude can invoke on demand.
At a glance
WHAT IT’S REALLY ABOUT
Build reusable Claude workflows using Skills, Cursor, and Claude Code
- Claire Vo explains Anthropic’s new Claude Skills feature as a lightweight way to package reusable, task-specific instructions (and optional files/scripts) that Claude can invoke on demand.
- She clarifies that a “Skill” is essentially a folder containing a required SKILL.md (with metadata + instructions) plus optional referenced resources and executable Python scripts for validation or consistent logic.
- After trying Claude’s built-in “create a skill” flow (which produced too many files and had download friction), she recommends a faster workflow: generate Skills in Cursor using the docs, then run and iterate them in Claude Code.
- She demos creating and using two practical Skills (changelog→newsletter and demo notes→follow-up email), then shows how to zip and upload a Skill to the Claude web UI, noting naming constraints (lowercase/hyphenated).
IDEAS WORTH REMEMBERING
5 ideasClaude Skills enable on-demand, reusable AI workflows.
Instead of copying prompts from docs or relying on always-on project context, Skills let Claude pull in task-specific instructions when the conversation context calls for them.
A Skill is a folder, not a “magic agent object.”
The core asset is a folder containing SKILL.md plus any supporting files; you either point Claude Code at that directory or zip it for upload to claude.ai.
Good Skills combine metadata + highly explicit instructions.
The YAML metadata (name/description) helps Claude choose when to use the Skill, while detailed instructions define the workflow, output format, and decision points.
Relative file references help manage context cleanly.
You can keep templates, examples, and extra guidance in adjacent files and link them (e.g., ./templates/...), so Claude only pulls what’s needed rather than bloating the main prompt.
Python scripts reduce variability and add guardrails.
Bundled scripts can validate formatting (e.g., YAML correctness, required sections) or implement consistent transformations, avoiding “LLM rewrites the tool every time” drift.
WORDS WORTH SAVING
5 quotesClaude Skills solves a really interesting problem for anybody using AI, which is reusable workflows with a set of instructions that you wanna call on demand.
— Claire Vo
A Claude Skill, the actual object, the thing that you make, is a folder.
— Claire Vo
My preference... is, you know what? If these models are so great at natural language, we should be able to define things in natural language.
— Claire Vo
It created 12 files when I really only needed five.
— Claire Vo
I just said, 'Okay, here's the latest changelog'... and it inferred what I wanted, based on the skills available in my folder.
— Claire Vo
QUESTIONS ANSWERED IN THIS EPISODE
5 questionsWhat heuristics does Claude use to decide when to invoke a Skill—metadata description only, keywords, or learned matching from examples?
Claire Vo explains Anthropic’s new Claude Skills feature as a lightweight way to package reusable, task-specific instructions (and optional files/scripts) that Claude can invoke on demand.
In your PRD Skill example, which parts are most essential (output format, questions, decision tree) versus “nice to have” for reliability?
She clarifies that a “Skill” is essentially a folder containing a required SKILL.md (with metadata + instructions) plus optional referenced resources and executable Python scripts for validation or consistent logic.
How do you prevent Skills from over-triggering when context is ambiguous (e.g., mentioning “changelog” in a different task)?
After trying Claude’s built-in “create a skill” flow (which produced too many files and had download friction), she recommends a faster workflow: generate Skills in Cursor using the docs, then run and iterate them in Claude Code.
What’s the best practice for organizing a larger “Skills repository” (naming, folder structure, versioning) once you have dozens of Skills?
She demos creating and using two practical Skills (changelog→newsletter and demo notes→follow-up email), then shows how to zip and upload a Skill to the Claude web UI, noting naming constraints (lowercase/hyphenated).
When does it make sense to move logic into a Python script versus keeping it purely in SKILL.md instructions?
Chapter Breakdown
Why Claude Skills matter: reusable, on-demand AI workflows
Claire introduces Claude Skills as Anthropic’s new way to package repeatable AI workflows that can be invoked across Claude Code, the API, and claude.ai. She frames the episode as a practical guide: what Skills are, how to create them, and where they fit into day-to-day product/engineering work.
Sponsor: ChatPRD (AI copilot for product docs and coordination)
A short sponsor segment explaining what ChatPRD does and who it’s for. Claire highlights integrations and the product’s positioning as a practical AI copilot for PM work.
What Claude Skills are (and how they differ from Projects/custom GPTs)
Claude Skills are described as task-specific instruction sets and context that can be called on demand. Claire contrasts them with Projects/custom GPTs, which tend to be static context tied to a workspace/chat rather than dynamically invoked, task-specific workflows.
Core benefits: discovery, context bundling, and consistent execution
Claire shares why Skills are useful in practice: they capture repeat tasks you’d otherwise copy/paste from prompt libraries. Skills can also bundle templates/examples via relative file references, and optionally include executable scripts for repeatable validation/processing.
What a Claude Skill actually is: a folder you zip or point Claude Code at
She demystifies the “asset” behind a Skill: it’s simply a folder containing a SKILL.md (instructions) plus optional adjacent files. You either keep it locally for Claude Code to use, or zip the folder and upload it to claude.ai.
Inside SKILL.md: metadata, instructions, and resource linking
Claire walks through the expected structure of a Skill file. The SKILL.md includes YAML-style metadata (name/description), detailed markdown instructions, and optional links to supporting files using relative paths.
Demo attempt: using Claude web to generate a Skill (helpful, but messy)
She tries Claude’s built-in “create skills” capability by prompting claude.ai to generate a PRD Skill. While the generated Skill is instructive (very detailed, includes decision trees and questions), the flow creates too many files and download friction makes it impractical.
Better workflow: create Skills in Cursor using the docs as source-of-truth
Claire shows a faster approach: use Cursor with Anthropic’s documentation link and generate a “Skill to create Skills.” Cursor produces a clean folder with SKILL.md, templates/examples, and even a validation script—quickly and with files already local.
Adding reliability: Python validation scripts for Skill structure and quality
Cursor’s generated Skill includes a Python validator that checks YAML/formatting, file existence, and basic content constraints. Claire notes this can be overkill, but it demonstrates how executable scripts can enforce consistency and reduce errors.
Testing in practice: invoking Skills with Claude Code (local repository)
With Skills stored locally, Claire runs Claude Code in the directory and asks it to use the Skill-creator Skill to generate a new workflow. Claude Code discovers available Skills, creates the new Skill folder, runs validation, and summarizes what it built.
Example Skill: changelog entries → user-facing newsletter
Claire demonstrates a concrete workflow: turning technical changelog entries into a subscriber-friendly newsletter. She highlights how Skills can be single-file and still powerful, and how Claude can infer Skill use based on context (e.g., the word “changelog”).
Example Skill: demo notes → personalized follow-up email
She repeats the pattern to quickly create another Skill that converts demo notes into tailored follow-up emails for trial prospects. The creation again takes only a few minutes and runs through the same validation and file-generation loop.
Uploading Skills to claude.ai: zip upload + naming constraints
Claire shows how to bring a local Skill into the Claude web UI by zipping the Skill folder and uploading it. She hits a constraint: skill names must be lowercase/hyphenated, fixes it, and confirms the Skill appears as an available capability in chat.
Wrap-up: recommended setup and next steps
Claire summarizes her recommended flow: maintain a local Skills repository, create a meta Skill that generates other Skills, and invoke them via Claude Code or upload zips to the web app. She closes with a call to subscribe/comment and teasers for future mini deep dives.
EVERY SPOKEN WORD
Install uListen for AI-powered chat & search across the full episode — Get Full Transcript
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome