Agent Skill Builder

Quickly scaffolds new Claude Code skills or translates ChatGPT projects into Claude Code format. This skill focuses on fast creation and correct structure. For iterative testing, benchmarking, and description optimization, use Anthropic's official skill-creator plugin.

Two Modes

Mode 1: Create New Skill (/agent-skill-builder new [name])

Scaffolds a fresh Claude Code skill with proper structure and best practices.

Mode 2: Translate ChatGPT Project (/agent-skill-builder translate [name])

Converts an existing ChatGPT project into a Claude Code skill.

Create New Skill

Step 1: Gather Intent

Ask the user (skip what's already clear from context):

1. Skill name: What should the /slash-command be called?
2. Purpose: What does this skill do?
3. Examples: How would this skill be used in practice?
4. Triggers: What would a user say that should invoke this skill?

Step 2: Plan Resources

Analyze each example to identify what should be bundled:

• Scripts (scripts/): Code rewritten repeatedly or needing deterministic reliability. Execute without loading into context.
• References (references/): Documentation Claude should consult while working. Loaded on demand.
• Assets (assets/): Files used in output (templates, images, fonts). Not loaded into context.

Step 3: Create Directory and SKILL.md

mkdir -p ~/.claude/skills/[skill-name]

Write SKILL.md following this template:

---
name: skill-name
description: "Describes what the skill does and specific triggers for when to use it. Write in third person. Include keywords users would say. Max 1024 characters. This is the PRIMARY triggering mechanism."
---

# Skill Name

[1-2 sentence description]

## Instructions

[Step-by-step guidance — use imperative form]

## Output Format

[Expected output structure]

## Constraints

[Limitations and rules]

Step 4: Add Reference Files (if needed)

For skills with multiple domains or variants, keep SKILL.md lean and split into references:

skill-name/
├── SKILL.md (overview + navigation, under 500 lines)
└── references/
    ├── variant-a.md
    └── variant-b.md

Reference from SKILL.md with clear guidance on when to read each file. Keep references one level deep.

Step 5: Test

1. Invoke directly: /skill-name
2. Let Claude auto-invoke by matching description keywords
3. Check that the skill triggers when expected and stays silent when not relevant

Translate ChatGPT Project to Claude Skill

Step 1: Extract ChatGPT Components

Request from user:

• System prompt/instructions
• Knowledge files (PDFs, docs)
• Conversation starters (optional)
• Actions/API configurations (if any)

Step 2: Map Components

Step 3: Convert Instructions

Step 4: Convert Knowledge Files

1. Convert PDFs/docs to Markdown and place in references/
2. For large files (>10k words), include grep search patterns in SKILL.md
3. For files >100 lines, add a table of contents at the top
4. Templates and assets go in assets/

Step 5: Write Frontmatter

The description field is the primary triggering mechanism. Write in third person and include both what the skill does and when to use it. Be slightly "pushy" with triggers to combat undertriggering.

Frontmatter Reference

Only name and description are required. Additional fields are Claude Code-specific extensions:

String Substitutions

Skill Locations

Writing Best Practices

These come from Anthropic's official skill authoring guide:

• Concise is key. Claude is already smart. Only add context Claude doesn't already have. Prefer concise examples over verbose explanations.
• Description is everything. Claude reads descriptions to decide which skill to use from potentially 100+ available skills. Be specific, include trigger keywords, write in third person.
• Progressive disclosure. Keep SKILL.md under 500 lines. Move detailed references to separate files. Claude loads them only when needed.
• Match freedom to fragility. High freedom (text instructions) for flexible tasks, low freedom (specific scripts) for fragile operations.
• Explain the why. Explain reasoning behind instructions rather than using heavy-handed MUSTs. Claude responds better to understanding motivation than rigid rules.
• Consistent terminology. Pick one term and use it throughout. Don't mix "API endpoint" / "URL" / "API route" / "path".
• No extraneous files. No README.md, CHANGELOG.md, INSTALLATION_GUIDE.md. Skills contain only what an AI agent needs to do the job.
• Forward slashes only. Use scripts/helper.py, never scripts\helper.py.

Prompt Skeleton (every skill and project prompt)

Every skill prompt and every ChatGPT or Claude project prompt uses the same five blocks, in this order:

1. Goal. One or two sentences on what the prompt produces and for whom.
2. Persona. Who the agent acts as, and how that person thinks about the work.
3. Chat protocol. How the conversation runs: one question per message, the stage name at the top of every message, a short explanation before each stage, and how to check the result after it.
4. Stages. Numbered stages. Each one ends with a stop point where the agent waits for the user's approval ("go") before the next stage starts.
5. Hard rules. The things the agent never does, written as plain rules.

Two more rules sit inside the skeleton:

• Capped revision rounds. Each stage allows one or two rounds of changes, then moves on. State the cap in the stage.
• No invention. The agent never invents facts, data, names or numbers. When something is missing, it marks the gap (for example [MISSING: price]) and asks, instead of filling it in.

When a handoff between stages is needed, have each stage write a markdown file the next stage reads.

After Creation Checklist

• SKILL.md created with name and description in frontmatter
• Description includes what AND when (triggers), written in third person
• SKILL.md under 500 lines
• Knowledge files converted to Markdown in references/
• References are one level deep from SKILL.md
• Large reference files (>100 lines) have table of contents
• Prompt follows the five-block skeleton, with stop points, capped revision rounds and the no-invention rule
• Skill tested with /skill-name and auto-invocation
• No extraneous documentation files