.tsx workflows by hand. You don’t memorize CLI verbs. You install one skill set into the agent you already use, then ask for an outcome (“add rate limiting and keep iterating until the tests pass”), and the agent reaches for Smithers, picks a ready-made recipe, and drives a durable, multi-step run on your behalf. You drive it through your agent.
Three steps. Each is one line your agent runs under the hood, then a sentence you say out loud.
Make every agent fluent in Smithers
Two commands wire Smithers into every coding agent on your machine (Claude Code, Codex, Cursor, Copilot, and ~20 more), no per-agent config by hand.The full per-agent matrix and this curated onboarding skill are covered in Installation and Agent Support.
skills add drops a skill for each Smithers command (ps, up, inspect, …) so the agent runs any verb without reading --help first; mcp add wires in the Smithers MCP server. Together they make the agent able to operate Smithers.For the mental model (when to reach for Smithers, the rule that the agent is an orchestrator that runs real work through Smithers rather than its own subagents, plus the full docs bundle it reads on demand), also install the curated smithers skill:
Initialize the repo so the ready-made workflows exist
From inside your project, scaffold the local workflow pack.This drops a
.smithers/ folder with task-specific and project-specific workflows already built: implement, plan, research-plan-implement, review, debug, ralph, and more. These are the recipes your agent will pick from. Nothing to author yet.Ask for an outcome, and watch the agent pick a recipe
This is the magic moment. You describe what you want done. The agent matches it to a workflow and starts a durable run.
You: Orchestrate an agent to add rate limiting and keep iterating until the tests pass.The agent reaches for the
implement workflow, wraps it in a loop that re-runs until your test command is green, and reports back when it lands. It survives a crash mid-run. Every finished step is persisted, so a restart resumes from the last completed frame instead of starting over.You: Give me an ordered plan for adding account-level billing limits. Don’t touch the code.The agent picks the
plan workflow: research and a structured, prioritized plan, no edits to your tree. Approve it, then ask it to implement.| You ask for | The agent reaches for | What you get |
|---|---|---|
| Build it and don’t stop until tests pass | implement (looped) | A durable run that iterates to green |
| A plan only, no code changes | plan | An ordered, prioritized plan to approve |
| Research, then plan, then build | research-plan-implement | The full chain, gated for your review |
Curious what happens under the hood? The Quickstart shows the
two commands your agent runs to launch a seeded workflow, and
Starters lists every plain-English recipe with the exact command
the agent runs for it. You stay in the conversation; the agent operates the
CLI.

Read next
Talk to Your Agent
The phrasing that gets the right workflow every time.
Watch and Steer
Follow a live run, clear approvals, and redirect without touching code.
Installation
The full per-agent matrix and the curated onboarding skill.