.tsx files, it’s phrasing. Say the outcome you want in plain language and your agent maps it to a Smithers workflow: render → execute → persist, durable, resumable, on any machine. Phrase it well and you get a high-quality result that runs while you’re away. Phrase it vaguely and you get a one-shot answer when you wanted a campaign.
One plain-English request, 'write a PRD', becomes a structured, durable workflow your agent runs on your behalf.
Name the outcome, not the steps
Describe the finished thing; your agent owns the route. You say what “done” looks like, and it reaches for the right recipe from the built-in pack.Set the finish line
Add a stopping condition and a one-shot becomes a loop. Say “don’t stop until X” and your agent iterates against persisted state until X holds (tests pass, the review signs off, a score clears a bar) or it hits an iteration cap.“Add account-level rate limiting, and keep iterating until the tests pass and the review approves.”Each pass: write code, run tests, send the diff to a review panel. It exits when tests are green and the synthesized review signs off, or burns the cap and reports back.
The cap matters. “Keep iterating until X” always carries a ceiling so a stubborn condition can’t loop forever. If the loop hits the cap without satisfying X, the run stops and surfaces what’s still failing, durable and inspectable, never silently spinning.
Ask for a gate
Say “check with me first” and you get a durable pause. The run suspends as a row in SQLite and costs nothing while it waits: no process burning, no tokens ticking, until you say the word.“Draft the release blog post, but don’t publish until I approve it.”Your agent writes the draft, then holds at a gate before publishing. Nothing ships, for a minute or a week, same cost either way.
Steer a live run
You redirect by talking to your agent, not by killing a process. Change your mind mid-run, say so, and your agent reaches for the right control verb:signal to nudge, cancel to stop a run, down to halt everything.
“Stop the kanban run and just focus on the billing ticket instead.”Your agent cancels the
kanban run and kicks off a focused implement run on the billing ticket. You never touched a run ID.
Approve from your seat
A pause is a question waiting for one word. Your agent shows you what’s on the table (the plan, the diff, the draft) and parks the run; because a suspended run is just a row, it waits indefinitely for free. Answer in plain language: “Approve.” “Ship it.” “No, redo the auth section first.” The run continues from exactly where it stopped, with no re-running the work already done.“approve”The gate clears, the run wakes, and the next stage executes against everything persisted. Deny instead and, depending on how the gate was set, the run stops, skips that step, or loops back.
A mission run renders its stages and gates as discrete, inspectable steps, so when it pauses for you, you see exactly what you're approving.
When not to hand it off
Don’t reach for Smithers when a single prompt already gets you there. One question with one answer, or a quick edit? Ask your agent directly. A workflow earns its keep when work spans stages, needs a finish line, needs your approval mid-flight, or should keep going while you’re away. On a one-liner it buys you nothing.
The same prompt-driven flow works across Claude Code, Codex, and more: any agent, any model, any machine.
Read next
- Watch and steer (how-to): follow a run once it is going, clear gates, and redirect it in flight.
- What you can do (reference): the full outcome catalog and the exact words that trigger each workflow.
- The few concepts you need (explanation): the vocabulary behind durable runs, gates, and loops.
- Agents and the CLI (under the hood): the models your workflows can drive and the verbs your agent runs for you.