Prompt optimizer
Content Guard analyzes screenshots from your devices. The accuracy of detection depends heavily on how you write your prompts. This guide explains how to write clear, effective prompts that produce reliable results.
How prompts work
Content Guard evaluates each of your active prompt-based items against the screenshot and returns a match/no-match result for each one. Your prompt acts as a content description to detect — Content Guard looks at the screenshot and decides: "Is the content described in this prompt visually present?"
Key behaviors:
- When uncertain, Content Guard leans toward "not detected" to minimize false alarms.
- Each prompt must be self-contained — they are evaluated independently.
- Content Guard pays special attention to "does NOT qualify" exclusion clauses in your prompt.
The golden rule: add exclusions
The single most impactful technique for reducing false positives is adding "does NOT qualify" exclusions to your prompt. After describing what you want to detect, explicitly state what should not trigger detection.
- What to detect — describe the visual content clearly.
- What qualifies — give concrete examples or thresholds.
- What does NOT qualify — list common false triggers.
Before and after
| ❌ Without exclusions (high false positives) | ✅ With exclusions (accurate) |
|---|---|
| "Food on screen" | "PHOTOGRAPHS or realistic IMAGES of actual food or drinks are visible: meals, snacks, beverages, coffee, desserts. A text-based menu listing dish names without food photographs does NOT qualify." |
| "Red screen" | "The entire screen is a solid red color with NOTHING else visible — no text, no menus, no icons. A red-themed advertisement or red background with content does NOT qualify." |
| "Advertisement" | "A specific sale offer with an explicit discount amount is visible: '50% OFF', 'SAVE $10', 'HALF PRICE'. Countdown timers, hashtags, or general signage without a specific discount do NOT qualify." |
Writing effective prompts
1. Describe visual appearance, not abstract concepts
Content Guard analyzes what is visible. Focus on what the content looks like, not what it means.
| ❌ Abstract | ✅ Visual |
|---|---|
| "Inappropriate content" | "Topless imagery, exposed body parts, explicit nudity" |
| "Error state" | "System error message with technical text such as crash dumps, blue screen of death, 'Application not responding', or stack traces" |
| "Political content" | "Political party logos, campaign slogans, election posters, or partisan symbols such as flags associated with political movements" |
2. Give examples and thresholds
Content Guard responds well to explicit examples, especially with numbers.
All caps text:
THREE OR MORE consecutive words in ALL UPPERCASE letters. Examples: "EMERGENCY EXIT ONLY" = 3 words = qualifies. "UP TO 70% OFF" = 4 words = qualifies. "POKER NIGHT" = 2 words = does NOT qualify. "DEPARTURES" = 1 word = does NOT qualify.
Brand detection:
The Coca-Cola logo, wordmark, or distinctive red-and-white branding is visible. The classic Coca-Cola script logo, a Coke bottle, or can with Coca-Cola branding qualifies. Generic red-and-white color schemes without the actual Coca-Cola brand do NOT qualify.
3. Distinguish photos from text
Content Guard can confuse text about something with an actual image of it. Be explicit.
Food and Beverages — PHOTOGRAPHS or realistic IMAGES of actual food or drinks are visible. Only actual food imagery (photos, illustrations of food items) qualifies. A text-based menu listing dish names like "Cheeseburger $8.99" with NO food photographs visible does NOT qualify — text names of food items are not food and beverages.
4. Be precise about screen states
For color screens, blank screens, or error states, define exact boundaries.
Blank empty screen — The screen shows ONLY a solid color or gradient with nothing else visible. Solid purple = blank. Solid blue = blank. A gradient with zero content = blank. If you see text, menus, images, dashboards, timers, logos, or any information = NOT blank.
Blue screen — The entire screen is a UNIFORM solid blue with NOTHING on it — no text, no menus, no logos, no icons, no error messages. If you can read ANY words or see ANY content on a blue background, it is NOT a blue screen.
Practical examples
Color/state detection
| Rule | Prompt |
|---|---|
| Red Screen | "A screen that is COMPLETELY and UNIFORMLY red with NOTHING else visible — no text, no menus, no countdown timers. If you can read ANY words or see ANY content besides red color, this is NOT a red screen." |
| Blank Screen | "The screen shows ONLY a solid color or gradient with nothing else visible. If text, menus, images, or any information is present, it is NOT blank." |
| White Screen | "The screen is entirely pure white with ZERO marks. If EVEN ONE character, number, symbol, or QR code is visible, it does NOT qualify." |
Brand & logo detection
| Rule | Prompt |
|---|---|
| Competitor Logos | "fast food brand logos or branding (e.g. McDonald's, Burger King, Wendy's, KFC). Generic restaurant imagery without identifiable brand logos does NOT qualify." |
| Coca-Cola Branding | "The Coca-Cola logo, wordmark, or distinctive red-and-white branding. Generic red-and-white color schemes without the actual Coca-Cola brand do NOT qualify." |
Content classification
| Rule | Prompt |
|---|---|
| Violence or Gore | "Explicit violence or gore: blood, fighting, physical assault, graphic injuries. The violence must be visually depicted. Text descriptions of violence or action movie posters without graphic content do NOT qualify." |
| Product Advertisement | "A specific sale offer with an explicit discount amount: '50% OFF', 'SAVE $10', 'HALF PRICE'. Countdown timers, hashtags (#SummerSale), or general signage without specific discount amounts do NOT qualify." |
| Food and Beverages | "PHOTOGRAPHS or realistic IMAGES of actual food or drinks. Text menus without food images do NOT qualify." |
Common pitfalls
| Pitfall | Problem | Fix |
|---|---|---|
| No exclusions | "QR code" triggers on any square pattern | Add: "A QR code (square matrix barcode with three corner squares). Map icons and decorative patterns do NOT qualify." |
| Too vague | "Text on screen" matches everything | Add threshold: "Dense paragraphs of body text covering most of the screen. Menus, price lists, or scattered labels do NOT qualify." |
| Instruction-style | "Check if there's a logo" | Description-style: "A company logo or brand mark is visible." |
| Ambiguous state | "Error screen" | Specific: "System error message: crash dumps, blue screen of death, 'Application not responding'. A colored background without error text does NOT qualify." |
| No threshold | "Lots of text" | Measurable: "THREE OR MORE consecutive words in ALL CAPS" |
Using the AI Helper
Not sure where to start? Use the AI Helper to generate a prompt automatically:
- Click "Add" → "Prompt" → click the "AI Helper" banner.
- Upload a sample image showing the correct (or incorrect) state.
- Choose a Positive prompt ("I want to see this") or Negative prompt ("I don't want to see this").
- Review the generated text and add exclusions — the AI Helper produces a good base description, but you should manually add "does NOT qualify" clauses for your specific use case.
The best prompts combine the AI Helper's description with your domain knowledge about what commonly triggers false alarms in your signage environment.