The prompt wizard writes an AI agent for you. You describe the agent in one line, and the wizard reads your company's website and a sample of your account: device types, organizations, devices, variables, their units and their usual ranges. It then asks only what it can't tell from that, writes the agent's system prompt, and saves the agent, ready to test.
To start, go to AI Center → Agents and click Create agent. The wizard has three steps: Your agent, Explore and Done.
Requirements
A plan that includes AI Center, with AI credits and an AI session available: drafting an agent uses both. See AI Center: Overview, Plans and AI Credits.
An account owner or admin user.
A free agent slot in your plan. The wizard saves the agent it drafts.
1. Describe your agent
On the Your agent step, under Describe your agent in one line, write who the agent helps and with what. Mention anything the wizard should know, such as where the data lives. For example: "Help shift supervisors see which machines are stopped and why. The data is in the Plant North organization."
Or click one of the common cases under Or start from a common case: Operations monitoring, Energy monitoring, Predictive maintenance, Building management, Cold chain, Tank levels, Yield monitoring, Air quality, Anomaly detection or Data Q&A. A case only fills in the description, so you can edit it before you go on.
Click Next.
To write the prompt yourself instead, click Start from a blank prompt. The New AI Agent page opens with an empty system prompt. See Create and Configure AI Agents.
2. Choose what to explore
On the Explore step, Website and account exploration, you tell the wizard where to look:
Your website: it's filled in from your profile when there's one. The wizard reads it so the prompt can describe who the agent works for. You can change it or leave it empty.
Explore my account's data: optionally, narrow the reading By device types, By organizations, By tags or By devices. A list with many values has a search box, and its N picked count clears that list's picks when you click it. Pick nothing to explore the whole account.
The wizard reads a sample, never every device: the device names, then the variables of about 10 of them. Your picks only steer that reading. They don't limit which devices the agent can use later, and they don't decide who sees which devices: that's set in each channel.
The summary at the bottom, We'll explore, shows what the wizard will read. Click Explore my account.
If the account has no devices yet, you can still draft the agent. The prompt stays general; draft it again once devices send data.
3. Answer the wizard's questions
The page splits in two.
On the left, Reading your account lists what the wizard reads as it goes, with your own names: for example, Listing devices in Plant North, Reading the variables of Compressor 3 and 9 more devices or Measuring 12 variables over the last 7 days.
On the right, the wizard asks one question at a time about what it couldn't tell from your account, such as thresholds, shifts or who reads the answers. Each question shows whether it's Recommended or Optional.
For each question you can:
Click one of the suggested answers.
Type your own answer, or ask the wizard to check something, such as "check device 00104", and click Send.
Click Skip.
Click Write the prompt now to stop the questions and get the prompt with what the wizard knows so far.
If the wizard couldn't fill some details from your account, it lists them before it finishes. Tell it the values, or leave one empty and it drops that line from the prompt. Then click Finish.
To change the description, the website or your picks, click Edit inputs. The conversation so far is dropped; your inputs are kept.
4. Try the agent
When the wizard is done, it shows Agent name is ready with the agent's description and Saved · not in any channel yet. The wizard saved it with the name, description, system prompt and four starter messages it wrote, the built-in Ubidots connector, the default model and zero data retention on.
Under Try it, click one of the test questions. The agent's page opens with that question sent in its chat preview.
Click See the prompt to read the system prompt the wizard wrote.
Click Open the agent to review its settings, attach connectors or change the model. See Create and Configure AI Agents.
The agent isn't available to anyone else until you add it to a channel. See Web Console for End Users and Talk to Your Industrial AI Agents through WhatsApp.
5. Draft an existing agent again
Open the agent and, on the Settings tab, click Prompt wizard next to the system prompt. Confirm with Draft again.
Drafting again replaces only the agent's system prompt, and attaches the Ubidots connector if the agent didn't have it. Its name, description and starter messages stay. If you edited the prompt by hand, copy anything you want to keep first.
If something goes wrong
"We couldn't finish reading your account." Click Try again. Your text and choices are kept.
"You're out of AI credits or AI sessions." Drafting uses AI credits and one AI session. Check your plan's limits, then try again.
"Your plan doesn't allow more agents." Click Copy the prompt so it isn't lost, delete an agent you don't use or upgrade, and try again.
"We couldn't save the drafted agent." Copy the prompt, then try again.
Website warnings: if the website can't be read, the reading log says why, for example "This address can't be read. Use your company's own website." or "The address only redirects elsewhere. Check it, for example with https://." The draft goes on without the website.
Frequently asked questions
Does the wizard change anything in my account?
Only the agent itself. The wizard reads your account and never creates, changes or deletes anything else in it. It reads with your permissions, so it only sees what you can see.
Does the wizard use AI credits?
Yes. Reading the account and writing the prompt use AI credits and count as one AI session, like a conversation in Chat.
Which websites does the wizard read?
Only your company's website (the one on the Explore step) and addresses you type in the conversation, plus their subdomains: the home page and a couple more pages. It doesn't read social networks or shared platforms such as LinkedIn, Facebook, YouTube or GitHub.
In which language is the prompt written?
The system prompt is written in English, which works best for the model. The agent's name, description, starter messages and the wizard's questions are in your language. To have the agent answer in another language, say so in the prompt or in your description.
Can I edit what the wizard wrote?
Yes. The system prompt is a normal prompt: open the agent and edit it on the Settings tab.
Related articles