Claude Skill Builder
Build Anthropic / Claude Agent Skills from VS Code.
Claude Skill Builder gives you a local workbench for creating Claude-compatible SKILL.md folders, drafting from free-form ideas, packaging skills for Claude.ai upload, validating skill structure, and syncing skills into Claude Code for local testing.
What It Creates
The extension creates standard Claude skill folders:
skills/
my-skill/
SKILL.md
references/
scripts/
assets/
evals/evals.json
It can also create authoring notes under:
workbench/
my-skill/
brief.md
authoring-prompt.md
Only skills/<skill-name>/ is the Claude skill deliverable. workbench/ is for planning, notes, and iteration.
Quick Start
- Open a workspace in VS Code.
- Open the Explorer sidebar.
- Find the
Claude Skills view.
- Click
Skill Builder: Interactive Builder.
- Describe the workflow you want as normal text.
- Click
Create Skill Draft.
The interactive builder creates a starter SKILL.md, starter evals, and a workbench brief from one form.
It can also import existing reference docs, scripts, and assets while the skill is being created.
Create A Skill From A Rough Idea
Use Skill Builder: Interactive Builder when you are still thinking out loud.
Good input looks like:
I need a skill that turns messy Jira security architecture updates into a weekly executive-ready status report. It should use our preferred sections, flag blockers, and create copy I can paste into Confluence.
Optional fields let you add:
- target surface: portable, Claude.ai, Claude Code, or Claude API
- trigger description
- should-trigger and should-not-trigger examples
- inputs Claude should gather
- workflow steps Claude should follow
- workflow modules that become ordered
references/01-*.md callout docs
- edge cases and decision points
- quality checks
- output standard
- examples, templates, and source materials
- references, assets, and progressive-disclosure plan
- reference files to create under
references/
- existing reference files to import with the VS Code file picker
- script plan
- script stubs to create under
scripts/
- existing script files to import with the VS Code file picker
- asset placeholders to create under
assets/README.md
- existing asset files to import with the VS Code file picker
- eval prompts
- target-surface notes
- external systems or MCP dependencies
- security and trust notes
When detailed fields are filled in, the extension also writes a workbench creation plan:
workbench/<skill-name>/creation-plan.md
That plan walks through trigger design, progressive disclosure, script rules, eval design, target-surface constraints, security, and the done checklist for a production-grade skill.
If you list reference or script files in the form, the extension creates starter files for them. Asset entries create notes in assets/README.md; use the file picker to import real templates, images, sample files, or other binary assets.
If you list workflow modules, the extension creates ordered module files:
references/01-intake.md
references/02-analysis-rules.md
references/03-output-format.md
SKILL.md is updated with ordered callouts so Claude knows when to read each module.
Add Existing Docs, Scripts, Or Assets
If you already have files, add them during Interactive Builder creation, from the Skill Editor, or by right-clicking a skill in the Claude Skills tree and choosing:
Skill Builder: Add Supporting Files
Then choose where the files belong:
references/ for Markdown docs, policies, schemas, examples, style guides, and other material Claude should read only when needed.scripts/ for deterministic helpers. Scripts should be non-interactive, support --help, use structured output when practical, and avoid hidden network or package-install assumptions.assets/ for templates, images, sample files, static tables, and other resources used by the skill.
The extension copies the selected files into the chosen folder and adds a small README.md note in that folder. After adding files, update SKILL.md so Claude knows when to read the reference, run the script, or use the asset.
Open The Skill Editor
Right-click a skill and choose:
Skill Builder: Open Skill Editor
From the command palette, the extension first lists skills from the configured workspace skills/ directory. It also includes a browse option so you can open any existing SKILL.md file or any folder that contains SKILL.md, even if it lives outside the workspace.
The editor opens a tabbed webview for:
- metadata: review and update the trigger description
- modules: inspect ordered
references/01-*.md modules or add a new module
- files: inspect and import
references/, scripts/, and assets/
- evals: inspect
evals/evals.json
- package: package ZIPs, sync to Claude Code, run validation, or copy a revision prompt
Package For Claude.ai
Right-click a skill in the Claude Skills tree and choose:
Skill Builder: Package Skill
The extension writes:
dist/<skill-name>.zip
The ZIP preserves the skill folder inside the archive, which is the structure Claude.ai expects for custom skill upload.
Sync To Claude Code
Right-click a skill and choose:
Skill Builder: Sync to Claude Code
By default, the extension copies the skill to:
~/.claude/skills/<skill-name>/
You can override the destination with:
skillBuilder.claudeCodeSkillsDirectory
After syncing, Claude Code can discover the skill from the filesystem.
Publish A Skill To Git
Git publishing is a manual, config-driven release step. Copy the example config first:
cp skill-builder.config.example.json skill-builder.config.json
Then set git.enabled to true and review the remote, branch prefix, validation, and include rules.
Right-click a skill and choose:
Skill Builder: Publish Skill to Git
The command opens a terminal and runs the cross-platform Node publisher. The publisher stages only skills/<skill-name>/ by default, can optionally include workbench/<skill-name>/ and dist/<skill-name>.zip, runs skills-ref validate when configured, shows the selected files, and requires a typed confirmation before commit and push.
You can also run it directly:
npm run publish-skill -- my-skill
Validate A Skill
Right-click a skill and choose:
Skill Builder: Validate Skill
This opens an integrated terminal and runs:
skills-ref validate "<skill-path>"
Install skills-ref separately if you want validation to run locally.
Health Check A Skill
Right-click a skill and choose:
Skill Builder: Health Check
The health check opens a Markdown report covering frontmatter, name/folder matching, description quality, SKILL.md length, workflow and quality-check sections, ordered module callouts, referenced supporting files, eval JSON, script CLI hints, and obvious secret-like values.
Commands
Skill Builder: Interactive Builder opens the free-form skill intake UI.Skill Builder: Open Skill Editor opens the tabbed metadata/modules/files/evals/package editor.Skill Builder: Create Skill creates a minimal starter skill through prompts.Skill Builder: Open Skill opens the selected SKILL.md.Skill Builder: Add Supporting Files copies existing docs, scripts, or assets into a skill.Skill Builder: Package Skill creates a Claude.ai upload ZIP.Skill Builder: Sync to Claude Code copies a skill to the Claude Code skills folder.Skill Builder: Publish Skill to Git commits and pushes a configured skill path through the manual publish script.Skill Builder: Validate Skill runs skills-ref validate.Skill Builder: Health Check opens a local structure and quality report.Skill Builder: Copy Free-Form Chat Prompt copies a prompt for deeper skill authoring work with an AI assistant.
Settings
skillBuilder.skillsDirectory: workspace-relative skill source directory. Default: skills.skillBuilder.distDirectory: workspace-relative ZIP output directory. Default: dist.skillBuilder.claudeCodeSkillsDirectory: sync destination for Claude Code. Default: ~/.claude/skills.skillBuilder.projectRoot: optional project root override.
Git publish defaults live in skill-builder.config.json, which is intentionally ignored by git. Use skill-builder.config.example.json as the starting point.
What This Extension Does Not Do
Claude Skill Builder does not upload skills into Claude.ai or the Claude API for you. It prepares the folder or ZIP artifact locally.
It also does not run Claude inside VS Code. The interactive builder creates a strong local draft and an optional authoring prompt, but deeper refinement still happens by editing the generated files or using your preferred AI assistant.
Security Notes
Treat skills like installable software. Review bundled scripts, dependencies, assets, and instructions that access external systems.
Do not hardcode secrets in a skill. For external services, prefer MCP-oriented instructions and document required tools and permissions clearly.
License
MIT License. See LICENSE.
Official References
Development
Install dependencies:
npm install
Compile:
npm run compile
Package:
npm run package
Publish a configured skill:
npm run publish-skill -- my-skill
