What Is the Safest Way to Use AI for Product Descriptions Without Publishing Invented Facts?

Summary

The steps

1. 1

   Create the drafting agent and the review agent

   Before you build the canvas, create two agents in Dave by voolama (hellodave.ai): one to draft the description and one to check it. Navigate to Agents in the sidebar. You need the Admin or Create role to see this section. The Agents section is not visible to users with the Use, Reporting, or Financial role.

   Drafting agent: Click New Agent, name it "Product Description Drafter", and click Create Agent. On the agent detail page, click Create Version. In the Config JSON editor, write a system prompt that instructs the agent to write from the facts supplied and nothing else. A starting config:

   {"systemPrompt": "You are a product copywriter for a small online store. Write a product description and three SEO meta-description options using ONLY the facts provided in the product brief below. Do not add specifications, materials, dimensions, or claims that are not in the brief. If a fact is missing, write around it rather than inventing it. Format: one paragraph description (80-120 words), then three meta-description options (each under 155 characters).", "model": "gpt-4o", "temperature": 0.4, "maxTokens": 600}

   The Config JSON fields Dave by voolama agents accept are: systemPrompt, model, temperature (0.0 deterministic to 1.0 creative), maxTokens, and topP. Click Create Version. The system assigns version number v1 automatically.

   Review agent: Create a second agent named "Product Listing Accuracy Checker". Click Create Version and write a review-focused system prompt:

   {"systemPrompt": "You are an accuracy reviewer for product listings. Compare the draft description below against the product facts provided. Return PASS if every claim in the draft is supported by the facts. Return FAIL followed by a list of unsupported claims if any claim in the draft goes beyond the facts. Be specific: name the exact phrase that is unsupported and the fact it contradicts or invents.", "model": "gpt-4o", "temperature": 0.1, "maxTokens": 400}

   Low temperature (0.1) makes the review agent consistent and predictable. Click Create Version.

   This two-agent design is the core difference between this workflow and a general approval workflow. A general approval workflow sends an AI draft to a human for review. This workflow sends the draft to a second AI first, which checks every claim against the facts before a human is involved. That sequence catches invented specifications automatically, rather than relying on the owner to catch them manually during approval.

   Do this

   • Go to Agents in the sidebar (Admin or Create role required).
   • Click New Agent, name it "Product Description Drafter", click Create Agent.
   • On the detail page, click Create Version. Paste the drafting Config JSON with your system prompt, model, temperature (0.4), and maxTokens (600). Click Create Version.
   • Click New Agent again, name it "Product Listing Accuracy Checker", click Create Agent.
   • Click Create Version. Paste the review Config JSON with temperature 0.1 and maxTokens 400. Click Create Version.
   • Note the version number for each agent (v1 for a first version).

   Example

   A hypothetical gift shop selling handmade ceramics might write the drafting agent's system prompt to include: "The store's tone is warm and personal. Mention the handmade origin only if it is stated in the brief. Never claim a specific clay type, firing temperature, or country of origin unless the brief states it." That specificity keeps the AI from inventing craft details that could mislead buyers or raise concerns under FTC endorsement guidelines (16 CFR Part 255).

   Best practice

   • Name agents after their function, not their model. "Product Description Drafter" is easier to find in a workflow than "GPT-4o Agent v3".
   • Set the drafting agent's temperature to 0.3 to 0.5 for readable, varied copy. Set the review agent's temperature to 0.1 or lower for consistent, predictable accuracy checks.
   • If you update your store's tone guidelines, create a new agent version with the updated system prompt and retire the old one. Retiring is a one-way action: you cannot un-retire a version, but existing running instances continue using the old version until they complete.
   • Keep the review agent's system prompt precise. Vague instructions like "check for accuracy" produce inconsistent results. Specific instructions like "return PASS or FAIL followed by the exact phrase that is unsupported" produce actionable output.
2. 2

   Create the workflow and open the Workflow Editor

   Navigate to Workflows in the sidebar. You need the Admin or Curate role to see this section. If the section is not visible, your account does not hold either role. See Why Can't I See Agents, Workflows, or Other Sections in Dave by voolama? for how to diagnose and fix your role assignment.

   Click New Workflow, name it "Product Description Drafting", and save. The workflow opens in the Workflow Editor: a three-panel canvas with the Node Palette on the left, the canvas in the center, and the Properties Panel on the right. A node and edge counter in the top-right corner of the canvas shows the current count as you build.

   Each time you click Save Version, the editor creates a new immutable snapshot of the graph stored as graphJson. Previous versions are preserved, so you can iterate without losing earlier designs. The editor does not auto-save: click Save Version after every meaningful change.

   Do this

   • Go to Workflows in the sidebar (Admin or Curate role required).
   • Click New Workflow, name it "Product Description Drafting", save.
   • The Workflow Editor opens with the Node Palette on the left.
   • Drag a Start node from the palette onto the canvas. Every workflow requires exactly one Start node.
   • Click Save Version to create your first snapshot.

   Example

   A hypothetical online kitchenware shop might name the workflow "Product Listing AI Draft" and add a description: "Staff enters product facts, AI drafts the description and SEO fields, an accuracy check runs automatically, and the owner approves before the listing goes live." That description appears in the workflow list and helps any team member understand the process without opening the editor.

   Best practice

   • Name the workflow after the business process, not the technology. "Product Description Drafting" is clearer than "Agent Workflow 2".
   • Add a description before you start building. It takes 30 seconds and prevents confusion when the workflow list grows.
   • Save a version after every meaningful change. Each save is non-destructive: the previous version is always available if you need to roll back.
3. 3

   Place the User Interaction and Information nodes

   The first two nodes after Start collect the product facts and keep your writing style guide visible on the canvas.

   User Interaction node (code: user_interaction, color: blue): Drag a User Interaction node from the Node Palette and drop it below the Start node. Click the node to open the Properties Panel. Change the label to "Enter Product Facts". Under the Fields section in the Properties Panel, configure the fields a staff member will fill in: product name (text), product category (text), key features (text), materials or ingredients (text), dimensions or size (text), and any restrictions or warnings (text). These fields become the workflow context that the drafting agent receives. Connect Start to this node by dragging from Start's source handle (bottom circle) to this node's target handle (top circle).

   When a staff member submits this task from the Task Inbox, the workflow resumes with those values in its context. The User Interaction node creates a task of type user_interaction in the Task Inbox. Only Admin and Use roles can see and complete tasks.

   Information node (code: information, color: indigo): Drag an Information node onto the canvas and place it beside the User Interaction node, not in the execution path. Label it "Writing Style Guide". Paste your store's tone and style guidelines into the node's config. The Information node does not affect execution and does not create a task. It keeps the style guide visible to anyone editing the workflow. The same guidelines belong in the drafting agent's system prompt from Step 1, so both stay in sync.

   Do this

   • Drag a User Interaction node below Start. Label it "Enter Product Facts".
   • In the Properties Panel under Fields, configure: product name, product category, key features, materials or ingredients, dimensions or size, restrictions or warnings.
   • Connect Start to User Interaction by dragging from Start's bottom handle to User Interaction's top handle.
   • Drag an Information node beside (not below) User Interaction. Label it "Writing Style Guide". Paste your tone and style guidelines into its config.
   • Click Save Version.

   Example

   A hypothetical outdoor gear shop might configure the User Interaction node with six fields: Product Name, Category (apparel, footwear, equipment, accessories), Key Features (comma-separated), Materials, Weight and Dimensions, and Safety Warnings. The Information node beside it holds the style guide: "Tone: practical and direct. Avoid superlatives like 'best' or 'ultimate' unless the brief states a specific award or certification. Keep descriptions under 120 words. Meta descriptions must be under 155 characters."

   Best practice

   • Keep the User Interaction fields to what the drafting agent actually needs. Every field you add is a field staff must fill in. Start with six and add more only if the agent's output is missing something important.
   • Use the Information node to keep the style guide on the canvas. When you update your brand guidelines, update both the Information node and the agent's system prompt at the same time so they stay in sync.
   • The Information node does not affect execution. Place it beside the flow, not in it, so the graph remains easy to read.
4. 4

   Add the Agent Interaction and Agent Review nodes

   After the staff member submits the product facts, the workflow passes them to the drafting agent, then immediately runs an automated accuracy check before any human sees the result. This is the sequence that makes the workflow accuracy-first rather than approval-first.

   Agent Interaction node (code: agent_interaction, color: purple): Drag an Agent Interaction node below the User Interaction node. Label it "Draft Description and SEO Fields". In the Properties Panel under Agent Assignment, select the "Product Description Drafter" agent and version v1. Connect the User Interaction node's source handle to this node's target handle. The agent receives the workflow context, including all the fields the staff member submitted, and produces a draft description and SEO meta-description options. The Agent Interaction node runs automatically and does not create a task.

   Agent Review node (code: agent_review, color: pink): Drag an Agent Review node below the Agent Interaction node. Label it "Check Claims Against Facts". In the Properties Panel under Agent Assignment, select the "Product Listing Accuracy Checker" agent and version v1. Connect the Agent Interaction node's source handle to this node's target handle. The Agent Review node runs an AI-powered automated quality check: it passes the draft description and the original product facts to the review agent, which returns PASS or FAIL with specific flags. Configure its output ports: one for PASS (connect to the Human Review node in Step 5) and one for FAIL (connect to a Routing node or back to the Agent Interaction node for a redraft).

   For a deeper explanation of why Agent Review comes before Human Review in this sequence, see How Do I Route AI-Generated Drafts Through Approvals and Compliance Checks in Dave by voolama?

   Do this

   • Drag an Agent Interaction node below User Interaction. Label it "Draft Description and SEO Fields".
   • In the Properties Panel under Agent Assignment, select "Product Description Drafter" and version v1.
   • Connect User Interaction to Agent Interaction.
   • Drag an Agent Review node below Agent Interaction. Label it "Check Claims Against Facts".
   • In the Properties Panel under Agent Assignment, select "Product Listing Accuracy Checker" and version v1.
   • Connect Agent Interaction to Agent Review.
   • In the Agent Review node's Properties Panel, configure two output ports: PASS (to Human Review) and FAIL (back to Agent Interaction or to a Routing node for manual inspection).
   • Click Save Version.

   Example

   A hypothetical electronics accessories shop might configure the FAIL path to route to a second Human Review node assigned to the store manager, rather than automatically redrafting. That way, if the AI accuracy checker flags a problem, a person decides whether to redraft or fix the facts before trying again. The PASS path goes straight to the owner's approval task.

   Best practice

   • Do not skip the Agent Review node. Sending an AI draft directly to Human Review means the owner is doing the accuracy check manually on every listing, which removes the time benefit of the workflow.
   • Set the review agent's temperature to 0.1 or lower. Higher temperature makes the accuracy check inconsistent: the same draft may pass one run and fail the next.
   • On the FAIL path, consider routing to a Human Review node rather than automatically redrafting. Automatic redrafting can loop if the agent keeps making the same mistake. A human break in the loop catches that pattern early.
   • The Agent Review node is an AI-powered check, not a legal guarantee. It catches claims that go beyond the supplied facts, but it does not verify that the facts themselves are correct. The staff member entering the product brief is responsible for the accuracy of the source data.
5. 5

   Add the Human Review node and End nodes

   The final gate before any description goes live is a Human Review task assigned to the owner or a designated approver. This is the step where a person reads the draft, the original facts, and the Agent Review result, and makes the final call.

   Human Review node (code: human_review, color: amber): Drag a Human Review node below the Agent Review node. Label it "Owner Approval". In the Properties Panel under User Assignment, assign it to the owner or the person responsible for approving listings. Connect the PASS output port of the Agent Review node to this node's target handle.

   When the workflow reaches this node, it creates a task of type human_review in the Task Inbox. The task payload contains the JSON data passed from the workflow: the original product facts, the AI-drafted description and SEO fields, and the Agent Review result. The owner opens the task, reads the Payload section, and clicks Approve, Reject, or Request Changes. Task status values are: pending (amber), in_progress (blue), completed (green), failed (red), cancelled (gray).

   For a detailed walkthrough of what the approver will see and how to complete a task, share How Do I Complete a Human Review Task in Dave by voolama? with your team before go-live.

   End nodes: Connect the Approve edge from the Human Review node to an End node labeled "Listing Approved". Connect the Reject edge to an End node labeled "Listing Declined". Connect the Request Changes edge back to the User Interaction node so staff can update the product facts and resubmit. Every path must end at an End node: dangling paths cause instances to hang at runtime.

   Click Save Version when the graph is complete.

   Do this

   • Drag a Human Review node below Agent Review. Label it "Owner Approval".
   • In the Properties Panel under User Assignment, assign it to the owner or designated approver.
   • Connect the PASS output port of Agent Review to Human Review's target handle.
   • Drag an End node and label it "Listing Approved". Connect the Approve edge from Human Review to it.
   • Drag a second End node and label it "Listing Declined". Connect the Reject edge from Human Review to it.
   • Connect the Request Changes edge from Human Review back to the User Interaction node so staff can update facts and resubmit.
   • Confirm every path leads to an End node. Check the node and edge counter in the canvas top-right.
   • Click Save Version.

   Example

   A hypothetical home goods shop might assign the Human Review node to the owner and configure the node's notification setting to "Custom audience: task assignee" so the owner gets an email when a new listing is ready for approval. The owner opens the Task Inbox, reads the draft and the Agent Review PASS result in the Payload section, and clicks Approve. The workflow reaches the "Listing Approved" End node and the instance completes.

   Best practice

   • Assign the Human Review node to a role rather than a specific user where possible. If the owner is unavailable, a role-based assignment lets any user holding that role pick up the task.
   • Add a comment when rejecting or requesting changes. The comment is stored in the audit log and helps the staff member understand what to correct.
   • Connect Request Changes back to the User Interaction node, not the Start node. This lets staff update only the facts that need changing, not restart the whole process.
   • Do not configure the workflow to auto-approve based on the Agent Review PASS result. The Agent Review checks that the draft matches the supplied facts; it does not check that the description is commercially appropriate, legally compliant with 16 CFR Part 255, or on-brand. That judgment belongs to the owner.
6. 6

   Activate the workflow and brief your team

   A workflow in Dave by voolama starts in Draft status and cannot be run by staff until you set it to Active. In the Workflows list, open the workflow and change its status to Active.

   Before you activate, run test instances in Draft status to confirm the two agents produce sensible output and the routing logic works. Submit a hypothetical product brief with complete facts to test the PASS path, and a brief designed to trigger the FAIL path, to confirm the Agent Review node catches the problem and routes correctly.

   Once active, brief two groups:

   • Staff who enter product facts: They will see a User Interaction task in their Task Inbox when a new product needs a description. They fill in the six fields and click Approve to submit. They do not need to understand the rest of the workflow.
   • The owner or approver: They will see a Human Review task in their Task Inbox when a draft has passed the automated accuracy check. The Payload section contains the product facts, the AI draft, and the Agent Review result. They read all of it, optionally add a comment, and click Approve, Reject, or Request Changes.

   Every decision is recorded in the audit log with the user's identity and a timestamp. The audit log is per-tenant and encrypted with AES-256-GCM authenticated encryption. For a practical checklist to use when reviewing AI-written listings before approval, see the companion resource at /resource/product-listing-review-checklist (awaiting publication alongside this guide).

   Disclosure: Dave by voolama is the product described in this guide, published by its maker, voolama LLC. This is vendor documentation, not independent legal or compliance advice.

   Do this

   • In the Workflows list, open the workflow and set status to Active.
   • Run a test instance with complete product facts (PASS path) and one designed to trigger the FAIL path.
   • Confirm the PASS test reaches the owner's Task Inbox and the FAIL test routes correctly.
   • Brief staff who enter facts: they complete User Interaction tasks in the Task Inbox.
   • Brief the owner: they complete Human Review tasks. Share the link to the Human Review task guide.
   • Confirm all team members have the correct role: Admin or Use to complete tasks, Admin or Curate to edit the workflow.

   Example

   A hypothetical craft supply shop might run two test instances before telling staff the workflow is live. The first test uses a complete brief for a set of watercolor paints: name, pigment types, paper compatibility, and pack size. The draft passes the accuracy check and lands in the owner's Task Inbox. The second test is designed so the review agent will flag a claim not in the brief, to confirm the FAIL path routes correctly. Both tests pass before the workflow goes Active.

   Best practice

   • Test the FAIL path deliberately. It is easy to confirm the PASS path works; the FAIL path is the one that matters for accuracy.
   • Keep the owner's approval step genuinely human. The AI drafts the description; the owner decides whether it goes live.
   • Check that every approver has the Admin or Use role. Users with other roles cannot see the Task Inbox. See Why Can't I See Agents, Workflows, or Other Sections in Dave by voolama? if someone reports they cannot find their tasks.

The accuracy-first node sequence at a glance

Who needs which role before you start

What this guide does not cover