How to Write Better Prompts for AI Visual Check
Overview
The prompt tells AI Visual Check what to look for in the scene. A strong prompt is clear, specific, and focused on one condition you want to detect.
Start simple. Add only the extra detail needed to improve accuracy.
Write a Simple Prompt First
Begin with a short instruction that clearly describes what should be checked.
- Basic Prompt
- Are the tables free of trash and food?
This gives the rule a clear objective without adding unnecessary detail.
Add Context to Improve Accuracy
If the rule is flagging items that should be ignored, add a short instruction to explain what belongs in the scene.
- Refined Prompt
- Are the tables free of trash and food? Ignore items placed in the center of the table, such as salt, pepper, ketchup, and sugar packets.
This keeps the rule focused on unwanted items while excluding objects that are expected to remain on the table.
Good Prompt vs. Poor Prompt
Use prompts that are clear and specific.
- Less Effective
- Check if the area looks bad.
- More Effective
- Are the tables free of trash and food? Ignore condiments placed in the center of the table.
The more effective example gives the rule a clear task and defines what should be ignored.
Use Alert Criteria to Control When Alerts Trigger
Once the prompt is working, add alert criteria to define when an alert should be sent.
- Example Alert Criteria
- Trigger the alert only when the table is not clean and no person is present.
This helps reduce unwanted alerts. For example, the rule can detect trash on a table but wait to alert until no person is nearby.
Test Before You Deploy
Use Test to confirm that the rule behaves as expected before rolling it out.
If the current scene does not include the condition you want to test, upload an image that does. This allows you to validate the rule against the exact scenario you want to detect.
Additional Tools
Reference Image
If the rule is still flagging items that should be ignored, use a reference image that shows what the scene should normally look like.
A reference image helps AI Visual Check compare the live scene to the expected scene and focus on what is out of place.
When to use a reference image:
- Expected objects in the scene are being flagged as problems
- The scene has fixed items that should be ignored
- The prompt alone is not enough to separate normal from abnormal conditions
NOTE: Using a reference image increases credit usage.
Analyzed Image Quality
If the rule is still not performing well, the issue may be image detail. Camera angle, distance, lighting, or scene setup can limit what the system can detect at standard quality.
Try the following:
- Start with Standard
- Move to High if more detail is needed
- Try Ultra if High still does not produce the expected result
NOTE: High and Ultra are often most effective when used with a reference image.
Credit Usage
Some settings can increase credit usage.
Credit usage may increase when you:
- Use a reference image
- Increase analyzed image quality to High or Ultra
NOTE: Use these options only when needed to improve results.
General Tips
- Keep prompts simple and specific
- Focus each rule on one main objective
- Add extra context only when needed
- Describe what should be checked, not every detail in the scene
- Small objects are harder to detect, especially with poor lighting, long distance, or low image detail
- Use testing to validate the rule before enabling it more broadly
If a specific object in the scene is causing false alerts, update the prompt to exclude it.
- Example
- If a nearby trash can is causing the alert to trigger, add: Ignore trash cans in the scene.
Best Practices
- Start with a simple prompt
- Refine the prompt only if needed
- Use alert criteria to control when alerts should be sent
- Use a reference image when expected objects are being treated as issues
- Increase image quality only when needed
- Test the rule before deployment
Troubleshooting
If the rule triggers too often
- Simplify the prompt
- Add ignore instructions for expected objects
- Use a reference image if the scene has fixed items that should not trigger alerts
- Review alert criteria to make sure the rule is not too broad
If the rule does not trigger when expected
- Test with a clearer example image
- Make sure the prompt clearly describes the condition you want to detect
- Increase image quality if the object is small or hard to see
- Check whether the camera angle or distance limits visibility
If small items are not detected
- Increase analyzed image quality to High or Ultra
- Use a closer or clearer camera view if available
- Add a reference image to improve scene understanding
If expected objects are flagged as problems
- Update the prompt to exclude those objects
- Use a reference image to show what normal looks like
Summary
For best results, use a prompt that is simple, specific, and focused on one task. Test the rule before deployment, then refine it only if needed. If expected objects are causing false alerts, use prompt exclusions or a reference image. If image detail is limited, try a higher analyzed image quality.