This article provides best practices for writing effective prompts in the Additional Context field when creating or editing an agent.
The Additional Context field gives Solink Agents extra instructions about what to look for, what to ignore, and how to interpret what it sees.
The more specific your instructions are, the more useful and consistent your agent results will be. Good prompts help agents understand your business environment and avoid false positives.
Note: Agents can support many different business types and use cases, including quick service restaurants, retail stores, cannabis dispensaries, music stores, convenience stores, and other monitored environments. The examples in this article are starting points that you can adapt to your locations, camera views, and operational goals.
What is Additional Context?
The Additional Context field is where you provide instructions that help the agent understand the situation it is monitoring.
Provide Additional Context to explain:
The business environment the agent is reviewing.
The specific activity, object, condition, or behavior you want the agent to detect.
What the agent should and should not look for.
Any visual details that are unique to your location.
How strict or cautious the agent should be.
Think of Additional Context as guidance you would give to a new employee who is reviewing camera footage for you. If a person would need extra details to make the right decision, the agent likely needs those details too.
How to Write Effective Agent Prompts
When writing Additional Context, use clear and specific language. Avoid short prompts that rely on the Agent to guess what you mean.
Follow the steps below to write effective prompts for your agents:
1. Describe the environment
Start by telling the agent what kind of area it is reviewing. This helps the agent understand what is normal for that scene.
For example:
"This camera shows the front counter of a quick service restaurant. Customers line up on the left side of the counter, employees work behind the counter, and the pickup shelf is on the right."
You can include details such as:
The type of business.
The camera view or area being monitored.
Where customers, employees, vehicles, products, or equipment are usually located.
What normal activity looks like in that area.
2. Be specific about the task
Clearly explain what you want the agent to look for. A broad prompt such as "look for problems" can lead to inconsistent results because the Agent has to decide what a problem means.
Instead, describe the exact condition you care about.
For example:
"Determine whether more than three customers are waiting in line and no employee is visible at the front counter."
This is better than:
"Tell me if the counter is busy."
3. Define what counts as a match
Tell the agent what should count as a positive result.
For example:
"Flag an event only when there are at least four customers standing in the checkout line and no cashier is present behind the register."
This helps the agent understand the threshold for when something matters.
4. Define what should be ignored
Include common situations that the agent should specifically ignore:
For example:
"Do not count employees restocking shelves as customers waiting for service. Do not count people who are walking past the counter but not standing in line."
This is especially useful for busy environments where many activities happen in the same camera view.
5. Include visual clues
Agents analyze camera footage, so providing visual clues is important.
For example:
"Customers waiting for service are usually standing in front of the register and facing the counter. Employees are usually behind the counter and wearing dark shirts with name tags."
Useful visual clues may include:
Uniform colours.
Store layout.
Product displays.
Queue direction.
Restricted areas.
Doors, counters, shelves, tables, or signage.
Objects that are often confused with the item you want detected.
6. Use steps for complex tasks
If the agent needs to reason through multiple details, write the prompt as a short sequence of steps.
For example:
Look for people standing at the front counter.
Determine whether they appear to be customers waiting for service.
Check whether an employee is visible behind the counter.
Only identify a service issue if customers are waiting and no employee is present.
Sequential instructions make the task easier to follow and can improve consistency.
7. Keep the prompt focused
Each agent should have a clear purpose. Avoid asking one agent to monitor too many unrelated things in the same prompt.
For example, avoid a prompt that asks the agent to check for:
Long lines.
Dirty tables.
Employees not wearing uniforms.
Open safe doors.
Missing products.
If these are all important, consider creating separate agents with focused prompts.
Prompt Structure
Use the following structure as a starting point:
Environment: Describe the business, camera view, and normal activity.
Task: Explain exactly what the agent should detect.
Match criteria: Define when the agent should consider the condition present.
Ignore: Explain what should not count.
Visual clues: Include details that help the agent tell the difference between similar situations.
Response guidance: If useful, explain the kind of result you want the agent to produce or prioritize.
Example:
"This camera shows the front counter of a quick service restaurant. Customers line up in front of the register, and employees work behind the counter wearing black shirts. Determine whether customers are waiting for service without an employee present. Count this as a match only when at least two customers are standing in line for more than a brief moment and no employee is visible behind the register. Do not count customers who are picking up completed orders from the shelf on the right. Do not count employees restocking or cleaning behind the counter as customers."
Examples of Strong Agent Prompts
The following examples show how to write Additional Context for different industries and scenarios.
Industry / Use Case | When to use it | Additional Context example | Why this works |
Quick service restaurant: Customers waiting without service | Use this when you want to know if customers are waiting at the counter and no employee is available to help them. | "This camera shows the front counter of a quick service restaurant. Customers line up in front of the register, while employees work behind the counter. Determine whether customers are waiting for service and no employee is present at the register. Count this as a match only when one or more customers are standing at the counter or in the line area and appear to be waiting to order. Do not count customers who are walking past the counter, picking up mobile orders from the pickup shelf, or already being helped by an employee. Employees usually stand behind the counter and wear black shirts or aprons." | Describes the camera view, explains the difference between ordering, pickup, and walking past, identifies what employees look like, and defines the condition that matters. |
Quick service restaurant: Dirty tables | Use this when you want to detect tables that need cleaning before another customer can use them. | "This camera shows the dining area of a quick service restaurant. Look for dining tables that are vacant but not ready for the next customer. A table should be considered dirty if it has visible food wrappers, cups, trays, napkins, spilled food, or other trash on top of it. Count this as a match only when the table is empty of customers and clearly has items that should be cleaned or removed. Do not count tables where customers are still seated. Do not count decorations, table numbers, condiment holders, or permanent fixtures as trash." | Defines what "dirty" means, separates vacant tables from occupied tables, and gives examples of trash and items to ignore. |
Retail store: Blocked aisle | Use this when you want to know whether customers may have difficulty moving through an aisle. | "This camera shows a retail store aisle. Determine whether the aisle is blocked or difficult for customers to walk through. Count this as a match when shopping carts, boxes, fallen merchandise, ladders, dollies, or other objects are blocking a significant part of the walking path. Do not count normal product displays or shelves that are outside the walking path. Do not count a person briefly standing in the aisle unless they are blocking the path for an extended period. The agent should focus on whether a customer could safely walk down the aisle." | Focuses on customer movement and safety, lists common obstruction types, and avoids triggering on normal displays or brief customer activity. |
Cannabis dispensary: Restricted area access | Use this when you want to monitor for people entering an area where only staff should be present. | "This camera shows the hallway leading to the secured inventory room in a cannabis dispensary. Only employees should enter the inventory room. Determine whether a person enters or attempts to enter the secured inventory room. Count this as a match when a person opens the inventory room door, walks through the doorway, or remains at the door while interacting with the handle or keypad. Do not count people walking past the hallway without stopping. Employees may wear store lanyards or uniforms, but still count the activity if someone enters the secured room because access should be reviewed." | Identifies the restricted area, explains the exact actions that matter, avoids false positives from people walking by, and makes clear that the event should be reviewed even if the person may be an employee. |
Music store: Guitar wall interaction | Use this when you want to monitor high-value merchandise. | "This camera shows the guitar wall in a music store. Guitars are displayed on wall hangers, and customers may browse in front of the display. Determine whether a person removes a guitar from the wall or handles a guitar for an extended period without an employee nearby. Count this as a match when a person takes a guitar off the wall hanger, carries it away from the display, or handles it for longer than a brief inspection while no employee is visible nearby. Do not count customers simply pointing at guitars or standing in front of the wall. Do not count employees adjusting the display or assisting customers." | Describes normal browsing, identifies the higher-risk activity, and distinguishes customer handling from employee assistance. |
Convenience store: Cooler door left open | Use this when you want to detect a possible equipment or product quality issue. | "This camera shows the beverage cooler doors in a convenience store. Determine whether any cooler door is left open after a customer or employee walks away. Count this as a match when a cooler door remains visibly open and no person is actively holding it or selecting a product. Do not count the door as open while someone is reaching into the cooler. Do not count reflections or bright lights on the glass as an open door. Focus on whether the door appears physically ajar after the interaction is finished." | Explains the timing of the issue, separates normal product selection from a door being left open, and warns the agent about reflections and lighting. |
Bad Prompt Examples
The following examples are too vague or too broad. Refer to these prompts as a guide for what to avoid.
Bad Prompt: Too Vague
"Check if the store is okay."
Why this is not effective:
"Okay" is not specific.
The agent does not know what condition to check.
The prompt does not explain what should count as a problem.
Better version:
"This camera shows the front checkout area of a retail store. Determine whether more than three customers are waiting in line and no cashier is visible at the register. Do not count customers browsing nearby shelves or walking past the checkout area."
Bad Prompt: Too Broad
"Look for safety issues, theft, customer problems, employee problems, and anything unusual."
Why this is not effective:
It asks the agent to monitor too many unrelated things.
"Anything unusual" can create inconsistent results.
The agent may focus on the wrong activity.
Better version:
"This camera shows the main retail aisle. Determine whether the walking path is blocked by boxes, carts, fallen merchandise, or equipment. Count this as a match only when the obstruction would make it difficult for customers to safely walk through the aisle."
Bad Prompt: Missing Context
"Tell me if someone is behind the counter."
Why this is not effective:
It does not explain whether being behind the counter is normal or unusual.
It does not identify who is allowed behind the counter.
It does not describe what the counter looks like.
Better version:
"This camera shows the service counter of a music store. Employees are allowed behind the counter, but customers should remain in front of it. Determine whether a customer enters the employee-only area behind the counter. Do not count employees helping customers or moving equipment behind the counter."
Bad Prompt: No Ignore Instructions
"Detect if the aisle is blocked."
Why this is not effective:
The agent may count normal customer activity as a blockage.
It does not explain what types of objects matter.
It does not define how much of the aisle must be blocked.
Better version:
"This camera shows a retail aisle. Count this as a blocked aisle only when boxes, carts, fallen merchandise, ladders, or equipment are blocking a significant part of the walking path. Do not count customers briefly standing in the aisle or normal product displays along the shelves."
Troubleshooting Agent Prompt Results
Certain prompts can cause potential issues with agent performance. See the Agent Troubleshooting guide for information on how to troubleshoot issues you may experience.
Prompt Checklist
Before saving your agent, review your Additional Context and ask:
Did I describe the camera view or business area?
Did I clearly explain the task?
Did I define what counts as a match?
Did I explain what should be ignored?
Did I include visual details that are unique to this location?
Did I avoid asking for too many unrelated things?
Would a new employee understand what to look for using only this prompt?