Step 1
Write down per task its objective, the connected systems with the permission granted to each, the actions allowed inside every system, the stopping conditions that end a run, and which person confirms each class of action.
The merchant wants an agent to act on its behalf but has not written down where that authority ends. In this case, the immediate task is to define an AI agent for bounded store tasks. The workflow should produce a task plan with an action log recording the objective, allowed actions and stopping limits behind each task.
First replace every [bracketed] placeholder with real store scope and constraints, then add any actual records or data you have, and paste it into your preferred AI assistant or Spark agent. Spark will follow this page's workflow to produce structured output and verification steps.
Act as a Shopify operations specialist.
Help me define an AI agent for bounded store tasks.
My store context is [describe the business, products and target audience].
My scope is [specify the affected products, pages, customers or reporting period]. Use these inputs: [the objective per task, each connected system with its granted permission, the actions allowed per system, the stopping conditions that end a run, and the named confirmer for each class of action]. Structure each task as evidence gathered from a named record, one proposed action inside the allowed set, and a verification read of the resulting state, pausing to log and request a decision whenever the next step needs a permission or record not granted.
Return the task plan with an action log confirmed against the merchant's stated operating rules for the specified task.
Show the relevant evidence or before-and-after examples and explain how I can verify the result.
If essential data is missing, ask for it and mark unsupported conclusions as unknown.
Do not invent metrics, product claims or customer facts.
For any publication, message, account or store change, use only connected tools within my explicitly authorized scope and report success only after verification; otherwise provide the draft or proposed steps.Split the work into input, judgment, and deliverable phases. Verify each step's output before moving to the next.
Write down per task its objective, the connected systems with the permission granted to each, the actions allowed inside every system, the stopping conditions that end a run, and which person confirms each class of action.
Structure each task as evidence gathered from a named record, one proposed action inside the allowed set, and a verification read of the resulting state, pausing to log and request a decision whenever the next step needs a permission or record not granted.
Deliver the task plan with an action log recording per step the objective, inputs read, action taken, verification result and stop reason, and review it against the merchant's stated operating rules.
Cover decision rules, frequent misunderstandings, and required inputs so the team aligns on prerequisites before execution.
Read-only or drafting work with an unambiguous success test, such as monitoring summaries, report drafts or tag suggestions, before anything writing to live settings, since verification is easier when the store state stays unchanged.
The run stops, the failure is written to the log with the input that caused it, and a decision is requested; blind retries risk duplicate writes, and guessing a missing input leaves the failure hidden.
Permissions are enumerated per system, any action outside them requires explicit confirmation and is logged with its approver, and the action log is reviewed on the merchant's stated cadence to catch scope drift early.
Start with scenarios under the same topic, then expand into cross-topic matches suggested by similarity.
Ciwi Spark builds Shopify-native AI tools and agents around the workflows merchants actually run: research, ads, translation, and customer support. Stop rewriting the same prompts by hand.