CloudRadial AutomationAI includes an AI Generate button in three places: the Designer (to generate a script), the Workflows page (to generate a whole workflow), and the Playbooks page (to generate a playbook). The same AI Generate label is used on the Custom tabs of the Agents and Extensions pages, covered in the articles on creating custom agents and building a custom extension. Each button turns a plain-language prompt into a working result, either from scratch or starting from an existing item, but the quality of that result depends heavily on what the AI can "see" in your workspace when it runs. This article explains, for each button, what it does, what context it relies on, and how to set yourself up for a good generation. It is written for workspace users who build automations.
- Generate a Script (Designer)
- Generate a Workflow (Workflows Page)
- Generate a Playbook (Playbooks Page)
- Starting From an Existing Item
- Model, Timing, and Limits
- Getting the Best Results
Generate a Script (Designer)
In the Designer, the AI Generate button in a script node's AI assistant turns a natural-language prompt into a single script in the language you choose — PowerShell, Python, or C#. You can also use it to refine a script you already have.
What It Relies On
When it generates a script, the AI is given:
- Your prompt
- Any installed Extensions you select — the AI reads each selected extension's usage guidance and the names of the secrets that extension needs to authenticate. Select every integration the script should use.
- The input variables available at that point in the workflow — the binding tokens from the trigger and from earlier steps (for example,
trigger.ticketIdornodes.fetch-tickets.output.results) - The current contents of the script, when you are refining an existing one
- The runner contract for the language you picked — the AI is taught exactly how a Runner reads inputs, loads and validates secrets, writes outputs, calls AI from inside the script, and reports logging and failures
Tips and Caveats
- If you select no Extensions, you get a generic script with no integration guidance
- Only reference Extensions you have actually installed. A made-up integration will fail at run time when the Runner cannot load its secret.
- Refine iteratively — generate a first draft, then ask for specific changes rather than rewriting the prompt from scratch
Generate a Workflow (Workflows Page)
On the Workflows page, the AI Generate button turns a prompt into a complete multi-step Workflow, plus a plain-language summary of what it built. Generation now happens in two steps: AutomationAI first expands your prompt against your installed Extensions so you can check and correct it, and then builds the Workflow from the text you approve. The AI both writes the Workflow and checks its own work before returning it. For a step-by-step walkthrough, see Generating a workflow with AI.
Step 1: Expand Prompt
After you enter a name and describe what the Workflow should do, select Expand prompt. AutomationAI keeps your text word for word and appends a section headed ## Resolved against your installed extensions, which spells out:
- The trigger
- Each step, with the Extension and tool that does it
- Each tool's required parameters, marked "not specified" when your prompt does not give a value
- The run-time inputs the Workflow needs
- Any assumptions the AI made where your prompt was ambiguous
Anything your installed Extensions cannot do is listed in a Not covered by your installed extensions warning and left out of the Workflow. Install or create an Extension that covers it, or generate the rest now.
The review step shows the result in an editable Expanded prompt box. Correct anything that is wrong, such as a "not specified" parameter you can fill in, then select AI Generate. AI Generate builds the Workflow from exactly the text in the box. Select Back to return to your original prompt instead.
If your prompt already says everything AI Generate needs, nothing is added. If the expansion fails, the review step still opens with your prompt as written, so you can edit it and generate anyway.
Step 2: Choose How the Steps Are Built
The Steps list sets how the Workflow does its work:
- Agent steps, one Agent node per step
- PowerShell scripts (the default), Python scripts, or C# scripts, where the language is applied to every script step
With Agent steps, each Agent node gets a goal that stands on its own (the objective, the exact values your prompt names, and the result to return), its allowed Extensions are limited to the ones that step uses, it is bound to a saved agent when your prompt names one, and a step that needs an earlier step's result receives it. The same Steps list appears in the Designer's AI Generation panel when you generate again.
What It Relies On
When it generates a Workflow, the AI is given:
- The prompt from the review step (your text plus the resolved section, as you edited it) and the Workflow name
- Your Steps choice: agent steps, or the script language applied to every script step
- The existing Workflow you picked under Start from, if any (see Starting From an Existing Item)
- Every Extension installed for your workspace — marketplace and custom — including each one's description, usage guidance, and required secrets
- Your available agents — both installed catalog agents and custom agents
- Your knowledge folders
- The full vocabulary of Workflow step types — start, PowerShell/Python/C# script, AI prompt, agent, knowledge write, knowledge consolidate, condition, for-each, and end — along with each step's options such as timeouts, retries, and parameters
- The data-binding patterns used to wire one step's output into the next step's input (for example,
{{ nodes.fetch-tickets.output.results }}) - The runner contract, so any script steps it writes are runnable
After drafting the Workflow, the AI runs a review-and-repair pass — it validates the result, and if it finds problems it repairs and re-checks, up to three rounds, before returning the finished Workflow.
Tips and Caveats
- The more relevant Extensions and agents you have installed, the better grounded the result
- Read the resolved section before you generate. Filling in "not specified" parameters and correcting a wrong tool choice there is faster than fixing the generated Workflow.
- Everything the AI references — every agent, Extension, and knowledge folder — must be real. Invented references either fail validation or fail when the Workflow runs.
- Only published Workflows are visible to the AI. Drafts are not included as building blocks.
- You can also reach this blade by converting an AI Playground run or conversation. See Converting an AI Playground run into a workflow.
Generate a Playbook (Playbooks Page)
On the Playbooks page, the AI Generate button turns a prompt into a complete, goal-directed Playbook — a sequence of "plays" that launch Workflows and run agents toward a goal. On success, the generated Playbook is saved as a Draft, exactly as if you had imported it.
What It Relies On
When it generates a Playbook, the AI is given:
- Your prompt, an optional home knowledge folder, and, when you start from an existing Playbook, that Playbook and the name you give the new one
- Your published Workflows, by name and description — drafts are excluded
- Your saved agents, by name and description
- Your knowledge folder names
- The Playbook structure — the Goal, Steps, Repeat, and Finish model; the play types (Workflow, agent, wait, approval, and reflect); per-play options such as what to do on failure and what to capture to knowledge; the template bindings for referencing earlier plays, the goal, the briefing, and the trigger; the repeat modes; and the finish rules
Like Workflow generation, Playbook generation runs a validate-and-repair loop of up to three rounds. If it succeeds, it saves a Draft Playbook. If it cannot produce a valid Playbook, it returns a clear error and saves nothing.
Tips and Caveats
- Publish the Workflows you want the Playbook to use before you generate. Draft Workflows are invisible to it.
- Every Workflow name and agent the Playbook references must match a real one exactly
- A home knowledge folder name must match exactly, or you can leave the Playbook company-wide
- The Name field appears only when you pick a Playbook under Start from
Starting From an Existing Item
Every AI Generate flow for Workflows, agents, Extensions, and Playbooks can start from one existing item of its own type instead of from scratch. That includes items you imported from the Marketplace and installed catalog agents and Extensions. You describe the changes you want, and AI Generate always saves a new item. The item you started from is never modified.
Picking a Source
Each AI Generate blade has an optional Start from list, set to None — start from scratch by default. Pick an item, and the prompt label changes to ask what should change in it. You can also start from a specific item directly with AI Generate from this…:
- On the Workflows page, in a workflow row's actions menu
- On the Agents page, on an installed agent's card or a custom agent's card
- On the Extensions page, on an installed Extension's card or a custom Extension's card
- In the Playbook builder header, for a saved Playbook
An installed agent or Extension opens AI Generate on the Custom tab, because the copy is saved as a custom item.
Naming the Copy
The name is filled in as <source> (copy), and you can change it. A Workflow copy is created under the name you enter. For a Workflow source, AI Generate starts from its newest draft, or from its latest published version when it has no draft.
Limits
- One source per generation
- A source can hold at most 128 KB of content. A larger one is refused before generation starts.
- If the AI cannot write out the whole copy in one response, the generation fails with a message such as "'<name>' is too large for AI Generate to copy in one response. Start from a smaller <kind>." Start from a smaller item instead.
- When a Workflow or agent source uses Extensions your workspace does not have, the generation summary lists them. Install them, or ask AI Generate to use something you do have.
Model, Timing, and Limits
All three buttons share the same underlying behavior:
- Generation uses your region's active AI model on Azure AI Foundry, run at a low temperature so the output is focused and consistent
- When you start a generation, the work runs in the background and the button reads Generating… and a progress line beneath it shows the current stage — grounding, generating, reviewing, and repairing — while the page polls for the result. A Workflow or Playbook that goes through repair rounds can take a minute or more.
- For a Workflow, Expand prompt reads Expanding… while AutomationAI checks your prompt against your installed Extensions
- Generation consumes AI usage, which counts toward your workspace's Usage Credits. The Expand prompt step is an AI call too and is recorded on your workspace's usage the same way.
- Content-safety filtering can occasionally block a request. For a Workflow, if a review or repair step is blocked, you receive the best result produced so far, with any remaining issues noted. For a script or a Playbook, a blocked request ends the generation with an error, and a Playbook is saved as a Draft only after it passes validation.
Getting the Best Results
The single biggest factor in generation quality is what real building blocks the AI can see. Before you generate:
- Install the Extensions the automation should use, so the AI knows the available integrations and their secrets
- Publish the Workflows and save the agents you want a Workflow or Playbook to build on
- Write a specific prompt — state the trigger, the systems involved, and the desired outcome
- When something close already exists, start from it with Start from and describe only the changes
- For a Workflow, check the expanded prompt and fill in anything marked "not specified" before you select AI Generate
- Review and test the result. Generated scripts, Workflows, and Playbooks are a strong starting point, not a substitute for verifying the automation before you rely on it.
Comments
0 comments
Please sign in to leave a comment.