# Universal Skill Builder > Autonomously design, research, construct, and validate new capabilities (Skills) for Gemini CLI, Codex CLI, and Claude Code. Use when creating a new skill, improving a SKILL.md file, scaffolding reusable skill folders, or deciding between architectures. Triggers on skill builder, SKILL.md, .agents/skills, skill triggers, and improve this skill. Use when this capability is needed. - Skill: `tomevault-io/universal-skill-builder` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add tomevault-io/universal-skill-builder` - Raw SKILL.md: https://api.skillmd.com/api/skills/tomevault-io/universal-skill-builder/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Research & Search - Author: tomevault-io (https://skillmd.com/u/tomevault-io) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/tomevault-io/universal-skill-builder --- # Universal Skill Builder Use this skill to build native skills that match current guidance for the specific target platform (Gemini, Codex, or Claude Code). **You MUST determine or ask the user which platform the skill is for before scaffolding.** ## Core Standard Default to repo-scoped skills in `.agents/skills/`. Build skills around one repeatable job. Keep the `description` exhaustive about when the skill should and should not trigger. Move deep material into `references/`, reusable templates into `assets/`, and deterministic helper logic into `scripts/`. ## Platform-Specific Knowledge Before scaffolding or modifying a skill, adapt your workflow to the specific platform requirements: 1. **Codex CLI**: - Heavily relies on `agents/openai.yaml` metadata files for discovery, UI metadata, and tool dependencies. - Skills start from metadata first. 2. **Claude Code**: - Skills are placed in `.claude/skills/` (or global `~/.claude/skills/`) or within plugins. - Claude Code relies on YAML frontmatter in `SKILL.md` (e.g., `name`, `description`, `allowed-tools`, `context: fork`). - Supports dynamic bash context injection using `` !`command` `` syntax directly in Markdown. 3. **Gemini CLI**: - Skills are typically located in `.gemini/skills/` or `.agents/skills/`. - Supports advanced `assistant.toml` command files with zero-latency injection (e.g., `@{file}` and `!{bash script}`). ## Decision Tree: Standard vs. Maestro Architecture Decide what architecture the skill needs: 1. **Standard Architecture** A repeatable workflow or specialized expertise for one clear job. Driven by a single `SKILL.md`. 2. **Maestro-Class Super Skill** If the skill requires complex terminal orchestration, external tool CLI usage, zero-latency dashboards, or extensive diagnostic checks. A Maestro skill leverages a multi-directory ecosystem (`references/`, `scripts/`, `templates/`), pre-computed bash scripts, auto-healing, and auto-backups. ## Authoring Workflow 1. Determine the target platform (Gemini, Codex, Claude Code) and architecture (Standard, Maestro). 2. Scaffold the skill: - Run `./.agents/skills/universal-skill-builder/scripts/scaffold-skill.sh --platform [--arch maestro]` 3. Write `SKILL.md` with sharp boundaries and imperative steps. Tailor the YAML frontmatter and formatting to the chosen platform. 4. Add `references/` only for material the AI should load on demand. 5. Add `scripts/` for deterministic helper logic (e.g., health checks, validations). 6. Validate the structure and trigger behavior. ## Validation Standard Every skill change should verify: - Structure is compatible with the target platform. - `name` and `description` exist in frontmatter. - Trigger wording is specific enough for implicit invocation. - The skill is scoped to one job. ## Assets ### Standard Assets - `assets/skill-template.md` - `assets/openai.yaml.template` - `assets/reference-template.md` ### Maestro Assets (Super Skill Factory) - `assets/maestro/SKILL.md.template` - `assets/maestro/command.toml.template` - `assets/maestro/doctor.sh.template` - `assets/maestro/auto-backup.sh.template` - `assets/maestro/log-action.sh.template` - `assets/maestro/KNOWLEDGE_MAP.md.template` --- > Source: [wtyler2505/ProtoPulse](https://github.com/wtyler2505/ProtoPulse) — distributed by [TomeVault](https://tomevault.io).