Workflow Setup — Productivity skill for Claude Code

workflow-setup

A Claude Code skill that designs and builds a Claude workflow for **your specific task**: one command researches the task on the web, picks a single-agent or multi-agent architecture, and, once you approve the plan, generates the project config (CLAUDE.md, subagents, skills, hooks, settings).

/workflow-setup "Build a weekly competitor-pricing report from 30 e-commerce sites"

Why

Configuring Claude Code for a new kind of work usually goes one of two ways:

• Too little. The defaults are used with no definition of done, so the agent stops when it thinks it's finished, reviews its own work leniently, and repeats the same mistakes.
• Too much. A pile of persona agents, MCP servers "just in case", and a 500-line CLAUDE.md. It burns tokens and context and isn't measurably better.

Getting it right takes research: what the domain's best practices are, which skills, plugins and MCP servers already exist and can be trusted, what tooling the environment needs, how to verify the output, and whether multi-agent actually pays off for this task. Anthropic's own data shows multi-agent setups use about 15× the tokens of chat. That pays off for broad, parallel research and is a bad trade for tightly coupled work like most coding.

`workflow-setup` automates that research and applies one rule throughout: **build the workflow around a verifiable definition of done, not around the number of agents.** Every component it adds has to justify its cost.

Pipeline

flowchart TD
    A["/workflow-setup "task""] --> B["Phase 0 · Intake
inspect project, OS, CLIs
≤3 clarifying questions"]
    B --> C{"Phase 1 · Parallel research
3–5 Sonnet subagents"}
    C --> D1["domain
literature & methods"]
    C --> D2["ecosystem
skills, plugins, MCP"]
    C --> D3["architecture
agent patterns & failures"]
    C --> D4["environment
tools, deps, OS limits"]
    C --> D5["verification