How to write plain-language workflow prompts that work
When you build a workflow by describing it, the description is the specification, and the same rules that make a good spec make a good prompt. The failure mode is not that the AI cannot build the workflow; it is that a one-line request leaves too much unsaid, so the platform has to guess at the parts you did not mention. The fix is a small amount of structure.
You do not need to write like an engineer. You need to name five things: what starts the workflow, which tools it touches, what should happen in order, where a human should approve, and how you will know it worked. Cover those and Ceven builds close to what you meant on the first pass, and the rest is refinement in the same plain language.
Name the trigger
Start with what kicks the workflow off. A new row in a sheet, an inbound email to a shared address, a form submission, a deal moving to a stage, a schedule like every weekday at 8am, or a manual run you kick off yourself. If you leave the trigger implicit, the platform has to infer it, and the inference is where surprises come from. One clear sentence about the trigger removes most of the ambiguity.
Name the tools by name
Say which systems are involved: read the lead from HubSpot, enrich it, draft in Gmail, log the result in a Google Sheet. Naming the tools does two things. It tells Ceven which connections the workflow needs, and it anchors each step to a concrete place data comes from or goes to. Vague nouns like the CRM or the database work, but named tools work better because they remove a round of guessing.
Describe the steps as an outcome, not as code
You are describing what should happen, not writing the logic. For each lead, research the company and the person, write a first-touch email that references something specific and true about them, and hold it for my approval reads clearly and gives the AI steps room to do the judgment work. Resist the urge to over-specify the parts the AI is good at, like phrasing, and be precise about the parts that are rules, like which field to write to.
Say where a human should approve
Call out the gate explicitly. Anything that sends to a customer, spends money, changes a record you care about, or publishes externally is a candidate. Hold the email for my review before it sends is one sentence and it changes the workflow from autonomous to supervised exactly where you want the supervision. If you do not name a gate, decide deliberately that the step is safe to run unattended rather than discovering it later.
State the definition of done
Tell the platform what a successful run looks like: the email is logged, the row is updated to sent, and I get a summary in Slack. A clear definition of done gives the workflow a place to end and gives you something to verify against on the first run. It also surfaces missing steps, because writing down done often reveals a handoff you forgot to mention.
Refine by running, not by rewriting
The first build will be close, not perfect. Run it once on real input and watch what each step does. Then correct in plain language: use the other email account, skip leads without a company, make the tone warmer. Refining against an actual run is faster than trying to specify everything up front, because you are correcting concrete behavior instead of imagining it.
Frequently asked
How long should a workflow prompt be?
As long as it takes to name the trigger, the tools, the steps, the gate, and the definition of done, and no longer. A few clear sentences usually beats a page, because extra detail on the parts the AI handles well adds noise.
What if I do not know which tools to name?
Describe the outcome and let Ceven propose the tools, then confirm or swap them. Naming tools is an accelerant, not a requirement. You can start from the outcome and fill in the systems as the platform suggests them.
Can I change the workflow after it is built?
Yes. You edit in the same plain language, and you can run it once to see the effect of a change before scheduling it. Refinement is the normal path, not a sign the first prompt was wrong.
Keep reading
What is agentic workflow automation
The older automation tools connect two apps with a rigid rule. Agentic automation takes an outcome you describe in plain language and figures out the steps.
How to add human approval gates to automations
The gate is what makes autonomy safe. Placed well, the AI does everything up to the decision and a person spends seconds approving. Placed badly, it is a bottleneck.
How to build an AI research workflow
Good research is repeatable, not heroic. Wrap it in a workflow that returns a cited brief every time, so the homework happens before the call instead of never.