Skill Creator

This skill guides through creating effective Agent Skills — markdown-based packages that teach AI agents specialized workflows, domain knowledge, and procedural capabilities.

Question Format (FR-UNIVERSAL.QA-FORMAT)

When asking the user a choice (IDE, scope, behavior decisions):

• Each question MUST be a numbered list item (1., 2., …) — not a heading, bold-only line, or paragraph.
• For multi-select questions, when the user delegates with agent's choice (or equivalent), pick the subset yourself, emit a one-line justification of the pick, and proceed without re-asking for confirmation.

About Skills

Skills are self-contained packages (directory with SKILL.md) that extend agent capabilities with:

1. Specialized workflows — multi-step procedures for specific domains
2. Tool integrations — instructions for working with specific file formats or APIs
3. Domain expertise — company-specific knowledge, schemas, business logic
4. Bundled resources — scripts, references, and assets for complex tasks

IDE Detection and Skill Placement

Skills work across multiple IDEs. Before creating a skill, determine the current environment and ask the user where to place it.

Control Primitives Map by IDE

Skill-Specific Paths

OpenCode also reads skills from fallback locations: .claude/skills/, .agents/skills/, ~/.claude/skills/, ~/.agents/skills/.

Detection Strategy

1. Check for IDE-specific markers in the project:
   • .cursor/ directory → Cursor
   • .claude/ directory → Claude Code
   • .opencode/ directory or opencode.json → OpenCode
2. If multiple detected or none → ask the user
3. Ask: personal skill (user-level) or project skill (shared via repo)?

IMPORTANT: Never create skills in ~/.cursor/skills-cursor/ — reserved for Cursor internals.

Core Principles

Concise is Key

Context window is shared with conversation, other skills, and user request. Every token competes for space.

Default assumption: the agent is already very smart. Only add context it doesn't already have.

Challenge each piece:

• "Does the agent really need this explanation?"
• "Does this paragraph justify its token cost?"

Prefer concise examples over verbose explanations.

Set Appropriate Degrees of Freedom

Anatomy of a Skill

Every skill: required SKILL.md + optional bundled resources in a directory:

skill-name/
├── SKILL.md              # Required - main instructions
├── references/           # Optional - detailed documentation
│   ├── api_reference.md
│   └── patterns.md
├── scripts/              # Optional - executable code
│   └── validate.py
└── assets/               # Optional - output resources
    └── template.md

SKILL.md (required)

• Frontmatter (YAML): name and description fields. These determine when the skill triggers.
• Body (Markdown): Instructions loaded AFTER skill triggers.

Bundled Resources (optional)

What NOT to Include

No README.md, CHANGELOG.md, INSTALLATION_GUIDE.md, or other auxiliary docs. The skill contains only what the agent needs to do the job.

Progressive Disclosure

Three-level loading to manage context:

1. Metadata (name + description) — always in context (~100 words)
2. SKILL.md body — when skill triggers (<5k words)
3. Bundled resources — as needed (unlimited)

Keep SKILL.md under 500 lines. Split content into separate files when approaching this limit. Reference them clearly from SKILL.md with guidance on when to read them.

Important: Keep references one level deep from SKILL.md. Avoid deeply nested chains.

Writing Effective Descriptions

Description is critical for skill discovery. The agent uses it to decide when to apply the skill.

Rules

1. Third person (injected into system prompt):

   • Good: "Processes Excel files and generates reports"
   • Bad: "I can help you process Excel files"

2. Specific with trigger terms:

   • Good: "Extract text/tables from PDF files, fill forms, merge documents. Use when working with PDF files or when the user mentions PDFs, forms, or document extraction."
   • Bad: "Helps with documents"

3. Include WHAT and WHEN (the WHEN phrase is MANDATORY, not optional):

   • WHAT: specific capabilities
   • WHEN: trigger scenarios — an explicit phrase such as "Use when …" or "when the user …"
   • Enforced: validate_skill.ts (and flowai's check-skills.ts for every skills/ primitive) reject a description that has no recognized WHEN-trigger phrase. Lazy forms ("How to X", "Helps with X", a bare domain label) fail.

Constraints

Common Patterns

Template Pattern

Provide output format templates when consistent structure matters:

## Report structure
ALWAYS use this template:
# [Title]
## Executive summary
[One-paragraph overview]
## Key findings
- Finding 1 with data
## Recommendations
1. Specific action

Examples Pattern

For output quality dependent on seeing examples:

## Commit message format
**Example 1:**
Input: Added user authentication with JWT tokens
Output: feat(auth): implement JWT-based authentication

Workflow Pattern

Break complex operations into steps with checklists:

## Workflow
- [ ] Step 1: Analyze input
- [ ] Step 2: Create mapping
- [ ] Step 3: Validate
- [ ] Step 4: Execute
- [ ] Step 5: Verify

Conditional Workflow Pattern

Guide through decision points:

1. Determine type:
   **Creating new?** -> "Creation workflow" below
   **Editing existing?** -> "Editing workflow" below

Feedback Loop Pattern

For quality-critical tasks:

1. Make edits
2. Validate: `deno run scripts/validate_skill.ts output/`
3. If fails -> fix -> validate again
4. Only proceed when validation passes

Anti-Patterns

• Windows paths: Use scripts/helper.py, not scripts\helper.py
• Too many options: Provide a default with escape hatch, not a buffet
• Time-sensitive info: Use "Current method" / "Old patterns (deprecated)" sections
• Inconsistent terminology: Pick one term, use throughout
• Vague names: code-review not helper or utils

Priming: Output Shape Teaches The Agent

Two lessons for authors changing a skill's output contract (what the final response should look like). Derived from debugging a skill whose output kept leaking into a file even after the write to file directive was removed.

Embedded examples are ground truth

When SKILL.md contains an example output block, the agent treats the shape of that example as the imitation target. A directive like "deliver findings inline in the response" competes with an example that starts with # Report Title (YYYY-MM-DD) + ## 1. Section + - [ ] item — and usually loses. That shape reads as "markdown file to save", and the agent saves it, regardless of the surrounding prose.

Rule: when changing an output contract, audit the example block, not just the directives. Drop file-shaped scaffolding:

• # / ## headings → plain-text category labels (1. Section Name).
• Timestamped titles (# Report (YYYY-MM-DD)) → remove; chat responses don't need titles.
• - [ ] checkbox syntax → plain - bullets; checkboxes imply "tracked task file".

The shape must match the target delivery medium. For inline chat output, the example should read like a chat message.

Remove the priming source, don't stack prohibitions

When unwanted behavior is primed by a rule, example shape, or system-level context (e.g., project AGENTS.md rules imported by a benchmark runner), the instinct is to add a negative constraint like Do NOT create files or Rule X does NOT apply to this skill. These rarely beat a shape-level or system-level prior — the agent has to pick a winner, often inconsistently.

Better approach: find the priming source inside the skill and remove it.

• Directive rule pointing at a file (Output Target: write to <path>) → delete the rule entirely.
• Section title with file connotations (Reporting, Save report, Persist) → rename (Deliver findings).
• Prose using "report" as a directive noun → switch to "findings" or "output".
• Example shaped like a document → reshape to chat-message form (see above).

Add explicit prohibitions only as a last resort, after reframing fails.

Skill Creation Process

Phase 1: Discovery

Gather from user:

1. Purpose and primary use case
2. Target IDE and storage location (personal vs project)
3. Trigger scenarios (when should agent apply this?)
4. Domain knowledge the agent lacks
5. Output format preferences
6. Existing patterns to follow

If context from prior conversation exists, infer the skill from discussed workflows/patterns.

Phase 2: Design

1. Draft skill name (lowercase, hyphens, max 64 chars)
2. Write specific, third-person description with WHAT + WHEN
3. Outline main sections
4. Identify supporting files needed (references, scripts, assets)

Phase 3: Implementation

1. Create directory structure
2. Write SKILL.md with frontmatter
3. Create supporting reference files
4. Create utility scripts if needed
5. Run deno run -A scripts/init_skill.ts for scaffolding if starting from scratch

Phase 4: Verification (BLOCKING gate)

Run validation and do NOT finish until it exits 0:

deno run -A scripts/validate_skill.ts <path/to/skill-directory>

validate_skill.ts is the portable per-skill FLOOR. It rejects a missing, oversized, or angle-bracketed description AND a description with no WHEN-trigger phrase (e.g. "Use when …"). If it fails, FIX the description and re-run — never proceed while it is red.

Two validators, two roles. validate_skill.ts (bundled here) is the portable floor that runs on ONE skill anywhere. Inside the flowai repo, scripts/check-skills.ts is the repo-wide gate wired into deno task check: it adds the catalog budget (name+description <100 tokens, ~400 chars — stricter than the floor's 1024-char description limit) and applies the same WHEN-trigger requirement to every skills/ primitive. A description can pass the floor yet fail the repo gate on the token budget, so when authoring inside flowai satisfy BOTH.

Description self-review rubric (BLOCKING — a "no" on any line means rewrite before finishing):

• WHAT — names the concrete capability (not the lazy "Helps with X" / "How to X")?
• WHEN — carries an explicit trigger phrase ("Use when …", "when the user …")?
• Specific — trigger terms a user would actually say, not a vague domain label?
• Third person — "Processes …", not "I can help you …"?
• Within budget — name+description ≲ 75 tokens (~300 chars) for headroom?

Checklist:

• validate_skill.ts exits 0 (line count, description present + WHAT + WHEN, no angle brackets)
• Description passes the self-review rubric above
• Consistent terminology
• File references one level deep
• No time-sensitive information
• No auxiliary docs (README, CHANGELOG)
• Examples are concrete, not abstract

Phase 5: Packaging

deno run -A scripts/package_skill.ts <path/to/skill-directory> [output-directory]

Creates a .skill zip file for distribution.

Phase 6: Iterate

1. Use skill on real tasks
2. Notice struggles or inefficiencies
3. Update SKILL.md or bundled resources
4. Test again

Design Patterns Reference

• Multi-step processes: See references/workflows.md
• Output formats and quality standards: See references/output-patterns.md