Meta Skill Generator

Turn a user's workflow brief into a concise, production-ready SKILL.md. Prefer shipping a minimal useful skill over an ornate one.

Default to English unless the user explicitly asks for another language.

Goal

Create or update skills that:

• trigger on the right requests
• encode only task-specific knowledge Codex would not already know
• use the lightest structure that still keeps execution reliable

Intake

Collect only the minimum inputs needed to design the skill:

• what job the skill should do
• what kinds of user requests should trigger it
• what output the skill should help produce
• whether the task needs scripts/, references/, or assets/
• any required tone, constraints, frameworks, or file formats

If the brief is incomplete, infer sensible defaults and note assumptions briefly after the draft instead of blocking.

Workflow

1. Distill the request into a one-sentence mission.
2. Choose the simplest structure that fits:

• workflow-based for stepwise procedures
• task-based for multiple operations
• reference-based for policies or specifications
• capability-based for multi-feature systems

1. Write frontmatter:

• name: short, stable, hyphen-case
• description: explicit trigger conditions; start with Use when...

1. Draft the body with only the sections the skill actually needs. Common sections:

• # Title
• ## Goal or ## Overview
• ## Intake
• ## Workflow or ## Task Guide
• ## Resource Use
• ## Output Contract
• ## Quality Bar

1. Add optional resources only when justified:

• scripts/ for deterministic or repeated automation
• references/ for large or variant-specific material
• assets/ for templates or files consumed by outputs

1. If the user asks for a full skill package, also add agents/openai.yaml with short UI text and a default prompt that explicitly mentions $skill-name.

Authoring Rules

• Keep the skill lean. Do not explain basics Codex already knows.
• Optimize for progressive disclosure. Put bulky details in references/ only if they are truly needed.
• Be concrete about triggers, inputs, outputs, and failure modes.
• Prefer direct instructions over theory.
• Do not create auxiliary docs such as README.md, CHANGELOG.md, or setup notes unless the user explicitly asks.
• Use examples only when they remove ambiguity.
• Match guidance strictness to task fragility:
• high freedom for broad creative work
• medium freedom for preferred patterns
• low freedom for fragile operations

Updating Existing Skills

• Preserve sections that already work well and only tighten what is vague, repetitive, or misleading.
• Improve the description first if the trigger conditions are weak.
• Remove generic filler before adding new sections.
• Do not expand scope unless the user clearly wants a broader skill.

Description Pattern

The description should answer:

• what the skill does
• when to use it
• which user phrases, tasks, or scenarios should trigger it
• what deliverable it helps produce

Preferred pattern:

yaml
description: Use when users ask to ...

Output Modes

When the user asks for only a SKILL.md, produce a complete SKILL.md and keep everything else optional.

When the user asks for a full skill package, produce:

• SKILL.md
• agents/openai.yaml if UI metadata would help
• only the minimal extra directories required by the workflow

Minimal Template

markdown
---
name: skill-name
description: Use when users ask to ...
---

# Skill Title

## Goal

State the job in 1-2 sentences.

## Intake

List only the inputs needed to do the job well.

## Workflow

Give the execution steps or decision rules.

## Resource Use

Explain when to read or run bundled files.

## Quality Bar

State the constraints, failure modes, and expected output quality.

Quality Check

Before finalizing, verify that:

• the name is short and stable
• the description would actually trigger the skill
• the body is concise and actionable
• optional resources are justified
• the draft matches the user's requested language and output scope