/bench-new-skill — Create New Skills

Scaffold new skills following established conventions, the official Claude Code skills spec, and Anthropic's Complete Guide to Building Skills.

When to Use

• User wants to create a new skill (anywhere — global, project-local, or bench-skills repo)

• User says "create a new skill", "add a skill", "new slash command"

• User wants to create a plugin (see references/plugin-guide.md)

Principles

• Start with use cases, not code. Define 2-3 concrete use cases before writing anything. See references/anthropic-guide-insights.md for the use case template.

• Description = [What it does] + [When to use it] + [Key capabilities] . Must include BOTH what and when. Under 1024 chars. No XML tags. Include specific trigger phrases users would say.

• Include only what Claude doesn't already know. Don't teach Claude how to write code — teach it your specific process, conventions, and standards.

• Be specific and actionable. Vague instructions like "validate the data" get ignored. Explicit instructions like "Run scripts/validate.py --input {file} " get followed.

• Self-contained skills. Each skill directory must work independently when installed.

• Progressive disclosure. Keep SKILL.md focused on core instructions (under 300 lines / 5,000 words). Move detailed docs to references/ and link to them.

• Right invocation model. Decide upfront: should the user invoke this, Claude invoke it, or both? Use disable-model-invocation and user-invocable accordingly.

• Support arguments when appropriate. If the skill operates on a target (file, PR number, issue), use $ARGUMENTS and add an argument-hint .

• No README.md inside skill folders. All documentation goes in SKILL.md or references/.

Process

Step 1: Understand the Skill

Ask the user:

AskUserQuestion: question: "What should this skill do? Give me 2-3 examples of what you'd say to trigger it."

From their response, identify:

• Use cases — Define 2-3 concrete scenarios (Use Case → Trigger → Steps → Result)

• Trigger phrases — What the user would say to invoke this

• Core process — What steps the skill should follow

• Output — What the skill produces

• Skill category — Document/Asset Creation, Workflow Automation, or MCP Enhancement (see references/anthropic-guide-insights.md)

• Agents needed — Does it need parallel subagents? For what?

Step 2: Choose Save Location

Ask the user where the skill should be saved:

AskUserQuestion: question: "Where should this skill be saved?" header: "Location" options: - label: "bench-skills repo (Recommended)" description: "Save to ~/bench-skills/skills/{name}/ — part of the shared skill collection, installable via npx skills add" - label: "Global personal" description: "Save to ~/.claude/skills/{name}/ — available in all your projects, personal use" - label: "Project-local" description: "Save to .claude/skills/{name}/ in the current project — scoped to this repo only" - label: "Plugin" description: "Package as a distributable plugin — for sharing across teams or community"

Use the chosen location for all file creation in later steps. If "Plugin" is selected, see references/plugin-guide.md.

Step 3: Name the Skill

Follow the {area}-{name} convention:

Area For

engineer-

Engineering workflow (plan, build, review)

product-

Product management (specs, requirements)

security-

Security auditing and best practices

knowledge-

Documentation and learning capture

bench-

Meta skills for bench-skills itself

Suggest a name and confirm with the user. Critical rules: kebab-case only, no spaces/underscores/capitals, no "claude" or "anthropic" in the name.

Step 4: Choose Content Type and Invocation Model

Content type — Determines what goes in SKILL.md:

• Reference content: Knowledge Claude applies to current work (conventions, patterns, style guides). Runs inline so Claude can use it alongside conversation context.

• Task content: Step-by-step instructions for a specific action (deployments, commits, code generation). Often invoked manually with /skill-name .

Invocation model — Who triggers the skill:

Setting User invokes Claude invokes Best for

Default (no flags) Yes Yes General-purpose skills

disable-model-invocation: true

Yes No Workflows with side effects (deploy, commit, send messages)

user-invocable: false

No Yes Background knowledge (legacy-system-context, coding-standards)

Ask the user which model fits their skill.

Step 5: Design Arguments and Dynamic Context

Arguments — If the skill operates on a target, use $ARGUMENTS :

• $ARGUMENTS — All arguments as a string

• $ARGUMENTS[0] , $ARGUMENTS[1] , or shorthand $0 , $1 — Positional access

• ${CLAUDE_SESSION_ID} — Current session ID (for logging/correlation)

• Add argument-hint: [issue-number] to show hints during autocomplete

Dynamic context — Use ! command`` to inject live data before Claude sees the prompt:

Context

• Current branch: !git branch --show-current
• Changed files: !git diff --name-only

The shell commands run first; their output replaces the placeholder.

Step 6: Decide on Execution Context

Inline (default) — Skill runs in the user's conversation. Best for reference content and skills that need conversation history.

Forked (context: fork ) — Skill runs in an isolated subagent. The skill content becomes the subagent's prompt (no access to conversation history). Best for:

• Tasks that produce a lot of output

• Skills that shouldn't see conversation context

• Read-only research (use agent: Explore )

• Planning tasks (use agent: Plan )

context: fork agent: Explore # or Plan, general-purpose, or a custom agent name

Only use context: fork when the skill contains an explicit task. If it's just guidelines/conventions, the subagent won't have an actionable prompt.

Step 7: Scaffold the Skill

Create the skill directory in the location chosen in Step 2:

bench-skills repo:

~/bench-skills/skills/{skill-name}/

Global personal:

~/.claude/skills/{skill-name}/

Project-local:

.claude/skills/{skill-name}/

Each skill directory contains:

{skill-name}/ ├── SKILL.md # Required — the main skill file └── references/ # Only if needed for large templates/catalogs └── {reference}.md

Write SKILL.md with this structure:

name: {skill-name} description: {triggering conditions — focus on when/why, not what} allowed-tools: [{tools needed}]

Optional fields — include only those that apply:

argument-hint: [{hint}]

disable-model-invocation: true

user-invocable: false

context: fork

agent: {agent-type}

model: {model-name}

/{skill-name} — {Short Title}

{One sentence explaining what this skill does.}

When to Use

• {Trigger condition 1}
• {Trigger condition 2}
• {Trigger condition 3}

Process

Step 1: {First Step}

{Instructions — use $ARGUMENTS where appropriate}

Step 2: {Second Step}

{Instructions}

...

Output

{What the skill produces and where it's saved.}

Next Steps

{What to do after running this skill. Link to other skills.}

Supporting files — If the skill needs large reference material, templates, or scripts:

• Keep SKILL.md focused on the essentials and navigation

• Put detailed content in references/ files

• Reference them from SKILL.md: See references/catalog.md for full details.

• Scripts go in scripts/ and can be in any language

Step 8: Validate

Check the skill against conventions and the Anthropic guide checklist:

• Name follows {area}-{name} pattern (kebab-case, no spaces/capitals, no "claude"/"anthropic")

• Description follows [What it does] + [When to use it] formula, under 1024 chars

• Description includes specific trigger phrases users would actually say

• No XML tags (< >) anywhere in frontmatter

• SKILL.md under 300 lines / 5,000 words (large content in references/ )

• No README.md inside skill folder

• Instructions are specific and actionable (not vague)

• Steps are numbered and actionable

• Error handling / troubleshooting included where appropriate

• Self-contained (no dependencies on shared files)

• allowed-tools includes only tools actually used

• Invocation model matches intent (disable-model-invocation / user-invocable )

• $ARGUMENTS used if skill operates on a target; argument-hint set

• If context: fork is used, skill contains an explicit task (not just guidelines)

• Frontmatter only includes fields that apply (see references/frontmatter-reference.md)

• Tested: triggers on obvious tasks, paraphrased requests, and does NOT trigger on unrelated topics

Step 9: Consider Plugin Packaging

Ask the user if this skill should be part of a plugin instead of a standalone skill. Use a plugin when:

• The skill needs to be shared with a team or community

• It should work across multiple projects

• It bundles multiple skills, agents, hooks, or MCP servers together

• It benefits from versioned releases

If yes, see references/plugin-guide.md for the plugin creation process.

If no, proceed with standalone installation.

Step 10: Install

Installation depends on the save location chosen in Step 2:

bench-skills repo: Create a symlink directly to the repo so edits are live:

ln -s /Users/elliottjacobs/bench-skills/skills/{skill-name} ~/.claude/skills/{skill-name}

Verify: ls -la ~/.claude/skills/{skill-name}

Global personal: Already available — saved directly to ~/.claude/skills/ .

Project-local: Already available — saved directly to .claude/skills/ .

Plugin: Test with claude --plugin-dir ./my-plugin . See references/plugin-guide.md.

After install, tell the user to start a new conversation and test with the trigger phrases identified in Step 1.

Skill Design Checklist

Before finalizing, verify:

• Did you define 2-3 concrete use cases? If you can't articulate specific scenarios, the skill isn't ready.

• Would this be triggered naturally? If the description requires users to know specific phrases, it's too narrow. If it triggers on everything, it's too broad. Test both directions.

• Is this distinct from existing skills? Check that it doesn't overlap with: engineer-plan, engineer-plan-review, engineer-work, engineer-review, product-brainstorm, product-prd, product-tech-spec, product-naming, security-audit, knowledge-compound.

• Is this teaching Claude something new? If Claude already knows how to do this without special instructions, the skill may not be needed. Skills capture how Claude should do it, not what Claude can do.

• Is this the right granularity? A skill should be a complete workflow, not a single step. But it shouldn't try to do everything either.

• Are the instructions specific and actionable? Vague instructions get ignored. Every step should tell Claude exactly what to do, not just what to consider.

• Is the invocation model right? Side-effect skills should be user-only. Background knowledge should be Claude-only. General skills can be both.

• Plugin or standalone? If sharing across projects/teams, consider a plugin.

Additional Resources

• For patterns, testing, troubleshooting, and success criteria from the official guide, see references/anthropic-guide-insights.md

• For complete frontmatter field catalog, see references/frontmatter-reference.md

• For plugin creation guidance, see references/plugin-guide.md

• Official docs: Skills | Plugins

• Anthropic guide: The Complete Guide to Building Skills for Claude