create-agnostic-skill

name	create-agnostic-skill
version	1.4.1
description	Use when adding a reusable workflow skill for AI coding agents. Scaffolds a new .agents/skills skill using the Agent Skills open standard.
argument-hint	[skill-name]
disable-model-invocation	true
allowed-tools	Read, Write, Bash, Glob, Grep, AskUserQuestion
user-invocable	true
metadata	{"internal":true}

Create Skill

Create a new skill for AI coding agents using the Agent Skills open standard. Skills live in .agents/skills/ (canonical source) and work across Claude Code, Cursor, Codex CLI, Gemini CLI, GitHub Copilot, and 20+ other compatible agents.

For deep-dive research on cross-provider compatibility, frontmatter behavior, and distribution patterns, see references/docs/skills-guide.md (bundled with this skill).

When to Use

Use when:

• Creating a new reusable workflow or capability
• Automating a repeated task that would benefit from structured instructions
• Adding a new command that should work across Claude Code, Cursor, Codex, and other agents
• Extending AI assistant capabilities for the repository

If you are creating a new oat-* skill:

• Prefer /create-oat-skill (it references this baseline guidance and adds OAT-specific conventions like {PROJECTS_ROOT} resolution and separator progress banners).

When NOT to Use

Don't create a skill when:

• The task is a one-time operation
• The workflow is still evolving rapidly
• The task is too simple to benefit from structured instructions
• Existing skills can be composed to achieve the goal

Arguments

Parse the following from $ARGUMENTS:

• skill-name: (required) Name for the skill in kebab-case (e.g., my-new-skill)

Workflow

Step 1: Gather Skill Information

If not provided in arguments, ask for:

Required:

• Skill name: kebab-case identifier (e.g., create-ticket, docs-review)
• Description: Prefer the formula Use when [trigger condition]. [What it does for disambiguation]. Allowed trigger stems: Use when, Run when, or Trigger when.
• Purpose: What problem does this skill solve?

Optional:

• Arguments: What parameters should the skill accept?
• Model invocation: Should the agent be able to invoke this automatically? (default: no)
• User invocable: (Claude Code only) Should this appear in the / menu? (default: yes)
• Tool restrictions: (Claude Code only) Which tools should the skill be allowed to use?
• Delegation: Will the skill dispatch subagents or require a fresh-context worker/reviewer? If yes, define the capability probe, authorization gate, fallback tier, and approval scope.
• Supporting files: Does the skill need templates, scripts, or reference files?

Step 2: Apply Progressive Disclosure

The context window is a shared resource. Skills share context with everything else the agent needs. Structure information in three levels:

1. Metadata (~100 words): Name + description in frontmatter, always loaded
2. SKILL.md body (<5k words): Loads when skill triggers
3. Bundled resources: Files in references/, scripts/, assets/ loaded as-needed

Writing principles:

• Challenge each piece of information—does the agent truly need this? Does this paragraph justify its token cost?
• Description is the primary trigger mechanism—include "when to use" there, not just in the body
• Prefer concise examples over verbose explanations
• Use imperative form (direct commands)
• Avoid duplication between SKILL.md and reference files

Step 3: Plan Skill Structure

Section structure (adjust based on complexity):

1. Title (H1) + brief description
2. When to Use / When NOT to Use (if applicable)
3. Arguments (if skill accepts arguments)
4. Workflow/Steps (use "Step 1, 2, 3..." naming)
5. Examples (both Basic Usage and Conversational)
6. Troubleshooting (if applicable)
7. Success Criteria

For detailed guidance, see references/skill-template.md.

Template:

---
name: skill-name
version: 1.0.0
description: Use when [trigger condition]. [What it does for disambiguation].
argument-hint: '[arg1] [--flag]'
disable-model-invocation: true
allowed-tools: Read, Write, Glob, Grep
user-invocable: true
---

# Skill Title

Brief description of what this skill does.

## When to Use

Use when:

- Condition 1
- Condition 2

## When NOT to Use

Don't use when:

- Condition 1
- Condition 2

## Arguments

Parse from `$ARGUMENTS`:

- **required-arg**: (required) Description
- **--optional-flag**: (optional) Description with default

## Workflow

### Step 1: First Step

Instructions...

### Step 2: Second Step

Instructions...

## Examples

### Basic Usage

\`\`\`
/skill-name arg1 --flag
\`\`\`

### Conversational

\`\`\`
Natural language request that triggers this skill
\`\`\`

## Troubleshooting

**Common issue:**

- Solution

## Success Criteria

- ✅ Criterion 1
- ✅ Criterion 2

Frontmatter notes:

• argument-hint, allowed-tools, user-invocable, context, hooks are Claude Code specific
• Other agents ignore unknown frontmatter fields, so it's safe to include Claude-specific fields everywhere
• name: max 64 chars for cross-provider portability (Codex allows 100, but 64 is the spec limit)
• version: include valid semver and start new skills at 1.0.0
• description: single line, ≤ 500 chars (Codex enforces single-line ≤ 500 chars; spec allows 1024)
• Bump version on future edits: patch for fixes/clarifications, minor for backward-compatible behavior additions, major for breaking workflow/interface changes

Writing the description field:

The description is your primary routing mechanism — agents load only name + description at startup across all installed skills, then semantic-match against the user's prompt. The SKILL.md body handles "what it does" once loaded. The description's job is to win the routing decision.

1. Lead with triggering conditions: "Use when…" / "Run this when…"
2. Include keywords for disambiguation: nouns + verbs that differentiate from similar skills
3. Don't summarize the workflow: providers route on description without reading the body
4. Front-load trigger keywords in the first 50 characters (truncation may occur at scale)
5. Use this default pattern: Use when [trigger condition]. [What it does for disambiguation]. Run when and Trigger when are also valid trigger stems when wording is clearer.

Examples:

• Bad: "Reviews code by checking spec compliance, then code quality, then creates PR"
• Good: "Use when reviewing code or checking PRs. Systematic quality and security analysis."

Present the plan and wait for user approval before creating files.

Step 4: Create Skill File

Create the skill at .agents/skills/{skill-name}/SKILL.md

Key requirements:

• Skill name matches directory name
• New skills include version: 1.0.0 in frontmatter
• Workflow steps use "Step 1, 2, 3..." naming
• Include both "Basic Usage" and "Conversational" example styles
• Use imperative form for instructions

Supporting files (create if needed):

Skills can include supporting files in subdirectories:

.agents/skills/{skill-name}/
├── SKILL.md                     # Main instructions (required)
├── scripts/                     # Executable code agents can run
│   └── setup.sh
├── references/                  # Additional docs loaded on demand
│   └── api-spec.md
└── assets/                      # Templates, images, data files
    └── template.md

Directory	Purpose
scripts/	Executable code that agents can run
references/	Additional documentation loaded on demand
assets/	Static resources like templates or data files

Step 5: Sync and Verify

After creating the skill, run OAT sync to update provider views:

oat sync

Verify:

• File created at .agents/skills/{skill-name}/SKILL.md
• Frontmatter syntax is valid
• Frontmatter includes valid semver version: (new skills start at 1.0.0)
• Skill appears in AGENTS.md
• Examples include both invocation styles
• If the skill name starts with oat-, run pnpm oat:validate-skills and fix any findings

Provide:

• Path to created file
• Summary of skill capabilities
• How to invoke: /skill-name
• Next steps (test the skill)

Best Practices

Naming

• Use kebab-case for skill names (e.g., create-pr-description, docs-new)
• Max 64 chars for cross-provider portability (lowercase letters, numbers, hyphens; no leading/trailing hyphen)
• Keep names short but descriptive
• Prefix related skills (e.g., docs-new, docs-update, docs-review)
• Skill directory name must match the name field in frontmatter

Content

• Context window is a public good—keep skills lean, challenge every paragraph
• Description is the trigger—include "when to use" in frontmatter, not just body
• Bump version for edits: patch = fixes/clarifications, minor = backward-compatible additions, major = breaking changes
• Keep SKILL.md under 500 lines / ~5,000 tokens (spec constraint)
• Use clear, task-based headings
• Include working examples for both invocation styles
• Document all arguments with defaults
• Keep command-like skills concise; complex workflows can be detailed
• Avoid duplication—info lives in SKILL.md or references, not both

Interactive Input

Skills that need user decisions (parameter choices, confirmations, disambiguation) should include interactive prompts in their workflow steps.

Write instructions portably: Use natural language like "Ask the user which approach they prefer" in workflow prose. All providers can handle this conversationally.

Host-specific structured input guidance: If the skill benefits from structured prompts, document the host split explicitly in the workflow:

• Claude Code: use AskUserQuestion when available
• Codex: use structured user-input tooling when available in the current Codex host/runtime
• Fallback: ask the same question in plain conversational text

Do not hard-code a specific Codex question tool name in skill prose unless the runtime contract is guaranteed. Prefer capability-based wording ("structured user-input tooling when available") so the skill remains portable across Codex hosts.

Claude Code enhancement: Add AskUserQuestion to allowed-tools in frontmatter. Claude Code renders these as structured UI prompts with selectable options, headers, and multi-select support. Other providers ignore the field and handle the same instructions as conversational questions.

When to include interactive input:

• Decisions with 2–4 discrete options (approach, config, scope)
• Confirmation gates before destructive or high-impact actions
• Collecting required parameters not provided in arguments

When NOT to:

• Long-running autonomous skills after work has started — resolve required decisions in a pre-work gate, then lock the choice for the run
• Questions with open-ended answers — just ask conversationally in the skill prose

Delegation and Subagent Capability

Skills that dispatch subagents, reviewers, workers, or fresh-context helper sessions must define a pre-work capability model. Do not let generated skills assume delegation is available, silently downgrade because authorization is required, or ask repeated approval questions mid-run.

When to include this section:

• The skill uses provider subagents, task workers, reviewers, or multi-agent execution
• The skill's quality depends on fresh context or isolated review/implementation
• The skill has a fallback mode when delegation is unavailable

Required guidance for delegation-capable skills:

1. Probe before work starts. Check whether the host can dispatch the needed worker/reviewer and whether dispatch requires user authorization.
2. Separate capability states. Report available, authorization required, or not resolved; do not treat authorization-required as unavailable.
3. Ask once when authorization is required. Use a concise question that names the delegation scope for the run. Approval should cover the stated worker roles and phases; a decline should select the documented fallback.
4. Lock the tier. Once selected, keep the execution mode for the run unless the user explicitly changes it.
5. Fail closed before side effects. If delegation is required for correctness and authorization is unresolved, stop before edits, writes, external side effects, or long-running work.
6. Document fallback behavior. If inline or single-agent execution is acceptable, explain when it applies. If it changes quality, artifact freshness, or review independence, say so.

Portable provider wording:

• Claude Code: use Task/subagent dispatch when available; if the skill benefits from structured prompts, include AskUserQuestion in allowed-tools.
• Cursor: use the provider's explicit agent invocation or natural-language agent handoff when supported.
• Codex: use multi-agent spawning when available; if the host requires explicit user authorization before spawning, ask before falling back.
• Fallback: use inline/single-agent execution only when the user declines delegation or the runtime truly cannot dispatch the required agent.

Good pre-work pattern:

### Step 0.5: Capability Detection

Before edits, writes, external side effects, or long-running work, detect whether the required helper agents are available.

- Available without authorization → Tier 1: delegated execution.
- Authorization required → ask once: "Authorize `{worker/reviewer}` delegation for this run?"
  - Approved → Tier 1.
  - Declined → Tier 2: documented fallback.
- Not resolved / unsupported → Tier 2: documented fallback.

Report:

`Selected: Tier {1|2} — {Delegated|Fallback}; Reason: {available|authorized|user declined delegation|dispatch unavailable|required role unresolved}`

Lock the selected tier for the run.

Anti-patterns:

• Selecting inline fallback just because the user did not pre-mention subagents, when the skill itself requires delegation and authorization can be requested
• Asking for authorization after implementation or review has already started
• Re-asking for each phase when the first approval explicitly covered the run
• Hiding fallback quality differences from the user

Progress Feedback

For multi-step skills, print brief progress updates so the user knows what's happening:

• Use [N/N] step indicators for sequential work (e.g., [1/3] Resolving dependencies…)
• For long-running operations (tests, builds, large diffs), print a start line and a completion line
• Keep it concise — don't print a line for every shell command

Shared References

• Skill-specific references go in the skill's own references/ directory
• Keep a shared doc's canonical copy in .agents/docs/ — edit it in one place, don't fork copies per skill
• If a distributed skill needs that shared doc at invocation time, vendor it into the skill's references/docs/ as a symlink to the canonical file: ln -s ../../../../docs/my-guide.md references/docs/my-guide.md The build (bundle-assets.sh, which copies with cp -RL) materializes the symlink into a real file, so the doc travels with the skill and the reference resolves wherever the skill is installed.
• Reference the bundled path from SKILL.md (references/docs/my-guide.md), not repo-root .agents/docs/. A bare .agents/docs/... reference only resolves inside this monorepo and dangles once the skill is installed in another repo.

Frontmatter Reference

Legend: ✅ supported | ⚠️ provider-specific | 💤 ignored | ❓ unknown

Field	Spec	Claude Code	Cursor	Codex CLI	Gemini CLI
name	✅ required	✅	✅	✅ required	✅
description	✅ required	✅	✅	✅ required	✅
license	✅ optional	❓	✅	💤	❓
compatibility	✅ optional	❓	✅	💤	❓
metadata	✅ optional	❓	✅	💤	❓
allowed-tools	⚠️ experimental	✅	❓	💤	❓
disable-model-invocation	❌	✅	✅	💤	❓
user-invocable	❌	✅	❓	💤	❓
argument-hint	❌	✅	❓	💤	❓
context / agent	❌	✅	❌	💤	❓
hooks	❌	✅	❌	💤	❓

Key takeaway: name + description are the only truly portable interface. Codex ignores unknown keys (safe to include Claude fields), so layer tool-specific fields on top of a portable baseline. For the full matrix, see references/docs/skills-guide.md (bundled with this skill).

Detail Level

Skill Type	Detail Level	Examples
Complex workflows	Detailed	docs-new, docs-review
Simple command-like	Concise	update-doc-refs, create-ticket
Reference/standards	Detailed	repo-documentation
Helper (auto-invoked)	Moderate	read-relevant-docs

Examples

Basic Usage

/create-agnostic-skill my-new-skill

/create-agnostic-skill deploy-preview

Conversational

Create a new skill called code-review that helps review pull requests

I need a skill for running database migrations

Reference

• Agent Skills Open Standard — the spec
• Claude Code Skills — Claude-specific features
• Cursor Skills — Cursor-specific features
• Codex CLI Skills — Codex-specific features
• Gemini CLI Skills — Gemini-specific features
• npx skills CLI — installing remote/community skills
• Skills best practices — authoring guidance
• references/docs/skills-guide.md — bundled deep-dive: compatibility matrix, resolved questions, patterns
• .agents/docs/reference-architecture.md — where skills/agents/docs live and why (OAT monorepo only; not bundled)

Troubleshooting

Skill not appearing in menu:

• Run oat sync to regenerate provider views
• Verify YAML frontmatter syntax is valid
• Check that skill name matches directory name
• Ensure user-invocable is not set to false (Claude Code)
• Restart the AI assistant

Skill invocation fails:

• Verify file created at .agents/skills/{name}/SKILL.md
• Check file path is correct
• Ensure no syntax errors in SKILL.md

Supporting files not loading:

• Verify files are in the correct subdirectory
• Check file paths in SKILL.md instructions

Success Criteria

Successful skill creation:

• ✅ Skill created at .agents/skills/{name}/SKILL.md
• ✅ oat sync run successfully
• ✅ For oat-* skills, pnpm oat:validate-skills passes
• ✅ Frontmatter valid
• ✅ Workflow uses "Step" naming
• ✅ Both example styles included
• ✅ Supporting files in skill directory (if applicable)
• ✅ Skill can be invoked with /skill-name