You don't have to design a workflow: this is the one Runway comes with. Every skill on this page is part of the Agent Runway plugin. The app installs it on first run, in Claude Code and Codex alike, so there is nothing else to add.
1 · Discover
Work out what you actually want before anything is built. This is where your judgment does the most good, so take as long as it needs.
mapAgent RunwayFor work too big for one session. It charts the decisions that have to be made, in the order they can be made, and keeps the map in your repo so any later session picks up where the last one stopped. It ends in a list of chunks, each small enough to spec.
grillAgent RunwayYour agent interviews you about the plan, one question at a time with its recommended answer, until every branch is settled. An unattended worker can't ask what you meant, so this is where those questions get answered.
2 · Prototype
When seeing beats reading, answer the question with something you can click.
prototypeAgent RunwayBuilds a throwaway prototype, often one HTML file or a few UI variants side by side, to settle a design question. It gets thrown away. The answer is what you keep.
3 · Spec
Write down what done looks like, in a form an unattended agent can check.
specAgent RunwayTurns the conversation into a spec in your repo, without starting a new interview: what it does, what it leaves out, its risk, and acceptance checks that say how to verify each one. Those checks become the tickets' checks.
4 · Tickets
This is the hand-off point. After your go, nobody needs to watch.
file-ticketsAgent RunwayProposes a short batch of tickets from the spec or the session, as one table. Each ticket has its title, its risk with a one-line reason, which agent should take it, what it waits on, where it runs, and one to three acceptance checks. Anything touching auth, payments, migrations, CI or secrets is never low risk. It files only after you type go, and a changed batch needs a fresh go.
5 · Build, check, review
The Runway app does this part, in plain code rather than an agent. Each ticket gets a fresh git worktree on a free Mac. The agent builds it, the repo's own check command runs, and a draft pull request opens. Then two fresh sessions review it on that exact commit, with none of the builder's context. The code review looks for bugs, risk and sensitive paths, and can raise the risk. The function review runs the check and walks every acceptance line, pass or fail. With one agent, both run on it. With both agents, the code review goes to the one that didn't build it. A failed review gets one fix pass. If a usage limit hits, the ticket waits in Up next until the limit resets, or goes to the other agent if you have both.
6 · Merge or ask
Low-risk tickets that pass review and checks, and touch no sensitive paths, merge on their own. Anything else becomes one question on your board, such as "Merge PR #2?", answered with a click or from a session.
runwayAgent RunwayAsk your agent what Runway is doing and what needs you, then say go, no, pause or resume. It acts only on your own words for that ticket. For a stuck ticket, it points you to take-back.
Take back
When a ticket stops on a failed review or check, finish it yourself instead of starting over. On the board, Take back · copy prompt copies one line. Paste it into Claude Code or Codex.
take-backAgent RunwayIt claims the ticket and stops its worker, keeping the branch and worktree. Then it reads the last run: what the worker said, both reviews' findings and the session itself. It tells you in a few lines what was tried and why it stopped, and you fix it together in the worktree. It always ends with one of three, and only with your OK. Give back sends it to Runway for fresh reviews, so a low-risk ticket can still merge itself. Done merges it on your typed go. Drop closes it. A ticket is never left taken back.
Getting better
The board shows how hands-off Runway really is: how many tickets merged without you, cycle time, cost per ticket, and where the rest stopped. When three tickets in a repo stop for the same reason within a week, Runway files a retro ticket to fix the cause. Its pull request links every run behind it, and a retro never merges on its own. With both agents, runway bench re-runs finished tickets on each one, so you can see which does better on your work.
Setup
setupAgent RunwayDoes the same checks as the app's first-run screen, from inside a session. It checks git, gh, tmux, Claude Code and Codex and their sign-ins, and walks you through each fix. Then it watches your first repo with its check command. You sign in yourself, and it never handles a password or token.
Why the line sits where it does
Everything above step 4 is judgment: what to build, what done means, and how risky it is. Runway leaves that to you, because a wrong spec builds the wrong thing very efficiently. Everything below step 4 is careful repetition, and Runway does it without you.
The shaping steps are inspired by Matt Pocock's open skills (mattpocock/skills), the way of working Runway was built around. Runway's versions are its own, written to hand work to unattended tickets.