Claude Code

Output is the first programming framework designed for AI coding agents — and Claude Code is how you use it

Output was designed for Claude Code from day one. We started with support for multiple coding assistants, then made a deliberate choice: ship only what we trust our own codebase with at GrowthX (we welcome community contributions for other assistants).

Every workflow keeps its files together — workflow.ts, steps.ts, types.ts, prompts, scenarios, traces, cost data — no scattering across folders. Describe what you want in plain English, and Claude Code writes correct, idiomatic Output code. The plugin gives it deep knowledge of the framework’s conventions, patterns, and common mistakes.

Installing the Plugin

When you create a new project with npx @outputai/cli init, the plugin is installed automatically. You’re ready to go.

To update it alongside your CLI:

npx output update --agents

If you ever need to install manually:

claude plugin marketplace add growthxai/output
claude plugin install outputai@outputai

What the Plugin Gives Claude Code

The Output plugin isn’t a wrapper around a few CLI commands. It’s a comprehensive system with 3 slash commands, 5 specialized agents, and 32 skills that give Claude Code expert-level knowledge of the Output framework.

When you start a Claude Code session in an Output project, the plugin loads automatically. It injects the full framework context — every convention, every pattern, every rule — so Claude Code understands Output from the first message.

Slash Commands

In practice, you rarely need to type these directly — Claude Code picks the right skill or command based on what you ask in plain English. But it’s useful to know what’s available so you understand what’s happening under the hood.

Command	What it does	Example
/output-plan-workflow	Produces a detailed implementation plan before writing code. Covers workflow structure, step boundaries, prompt design, error handling, and testing strategy. Saves to .outputai/plans/.	A workflow that researches a company and produces a sales brief
/output-build-workflow	Takes a plan and generates working code — workflow.ts, steps.ts, types.ts, prompts, evaluators, and test scenarios. Every file follows Output conventions.	.outputai/plans/2026_03_19_company_research/PLAN.md
/output-debug-workflow	Checks infrastructure, pulls execution traces, identifies the failure, and suggests a fix. Systematic — no guessing.	The company_research workflow is timing out on the scrape step

Specialized Agents

The plugin includes 5 expert agents that Claude Code delegates to automatically based on what you’re asking.

Agent	Expertise
Workflow Planner	Designs workflow architecture, step boundaries, orchestration patterns
Workflow Quality	Reviews code for framework compliance, catches common mistakes
Workflow Debugger	Diagnoses execution failures using traces and error patterns
Prompt Writer	Creates and reviews .prompt files — Liquid.js templates, YAML frontmatter, provider config
Context Fetcher	Retrieves existing patterns from your project to ensure consistency

Skills

Skills are the building blocks that power the commands and agents — 32 in total across five categories:

• Development — Creates every file type in an Output project with correct patterns, imports, and conventions
• Workflow operations — Run, start, check status, retrieve results, list workflows
• Monitoring and debugging — Analyze traces, check service health, identify failure patterns
• Error diagnosis — Six skills for the most common mistakes: wrong zod imports, non-determinism in workflows, try-catch swallowing retries, missing schemas, direct I/O in workflow functions, using axios/fetch instead of @outputai/http
• Credentials — Initialize encrypted stores, edit secrets, wire to environment variables

What Gets Enforced

The plugin enforces Output’s conventions so you don’t accidentally create patterns that break in production:

• Import z from @outputai/core, never from zod directly
• All I/O goes in steps, never in workflow functions
• Use @outputai/http for HTTP calls, not axios or fetch
• Use @outputai/llm for LLM calls, not provider libraries directly
• ES module imports always use .js extensions
• Every step and workflow has input and output schemas
• Prompt templates use Liquid.js syntax, not Handlebars
• Don’t wrap step calls in try-catch (it swallows the retry logic)

These aren’t arbitrary rules. Each one exists because we hit the bug in production at GrowthX and built the guardrail so you don’t have to.

Want to dive deeper?

You can see all skills, agents, commands, and hooks here in our repo.

What’s Next

Learning Path

Learning Path. A structured guide through the rest of the documentation.

Workflows

Workflows. Deep dive into workflow structure, patterns, and orchestration.