This section provides step-by-step guides showing exactly how to configure different types of annotation projects. Follow these walkthroughs to understand what the platform can do and how to reproduce these examples yourself.

🚀 30-60 minutes → first project → first results

Choose an example below and follow the instructions.

 

Choose your example

📸 Query selection for image annotation

⏱️ 30-60 min | 💰 ~$3-4 per task| 👥 3000+ experts

Experts evaluate two search queries and select the best match for each image

 

🔍 Search query generation

⏱️ 30-60 min | 💰 ~$3-4 per task| 👥 3000+ experts

Generate 2 distinct search queries for each image

 

🖼️ Image classification

⏱️ 30-60 min | 💰 ~$3-4 per task| 👥 3000+ experts

Classify images into predefined categories

 

🏞 Images comparison

⏱️ 30-60 min | 💰 ~$1 per task| 👥 3400+ experts available

Evaluate two suggested images and select the one that best matches specific criteria

 

Use Case: Query selection for image annotation

Experts evaluate two suggested search queries for each image and select the one that best matches the image. The selected query should be optimized for searching within a limited dataset with diverse images of similar types but no exact duplicates.

⏱️ 30-60 min | 💰 ~$3-4 per task | 👥 3400+ experts available

 

Step 1: Prepare your dataset

Actions:

  1. Create a JSON file with images, metadata, and two suggested queries per image.
  2. Save as images_with_queries.json.
  3. Ensure image URLs are publicly accessible.

Or use example files:

Your dataset structure (images_with_queries.json)

[
  {
    "image_id": "IMG001",
    "image_link": "https://example.com/images/chart_bullying_2015.jpg",
    "author": "Research Team",
    "date_created": "2015-03-15",
    "title": "Survey Data on Bullying",
    "suggested_query_1": "Find a chart of children being bullied in school in 2015",
    "suggested_query_2": "Graph showing bullying prevalence by age group in 2015"
  },
  {
    "image_id": "IMG002",
    "image_link": "https://example.com/images/diagram_process_flow.png",
    "author": "Design Team",
    "date_created": "2023-06-20",
    "title": "Customer Onboarding Process",
    "suggested_query_1": "Display the workflow diagram showing the customer onboarding process",
    "suggested_query_2": "Show me a process flow chart"
  }
]

Check: Valid JSON, all fields filled, image URLs accessible

Step 2: Create project and describe your task

Step 2: Create project and describe your task

Actions:

  1. Log into the platform.
  2. Create Project Space (e.g., "My First Projects").
  3. Click "Create New Project" to Create Project within a Project Space.

    image (16).png
  4. Describe your task to the agent.

What to type:

I need experts to evaluate two suggested search queries for each image and
select the one that best matches the image. The dataset contains diverse
images of similar types and topics, but no exact duplicates. The selected
query should be optimized for searching within this limited dataset.

Selection criteria:
- Relevant context within dataset (sufficient detail to distinguish from
  similar images, but not overly specific)
- Natural and grammatically correct phrasing (like a realistic user request)
- Avoid vagueness (not too broad)
- Avoid overspecificity (not excessive details)
- Dataset-aware specificity (reasonable detail is enough for accurate retrieval)

For each image, experts should select one query and provide a brief reason
for their selection.

Or use description from: Project Description File

On the right side, you'll chat with the AI Assistant. After each step, you'll be invited to Accept or modify the generated results shown in the main window on the left. Once you Accept, you'll see the agent's report on the completed workflow. To view the next step's results, you must also accept the configured content.

image (17).png

 

What you check: ✅

  • The project goal mentions SELECTING the best query (not generating).

  • The annotation task emphasizes evaluation and selection.

  • Selection criteria include relevance, natural phrasing, appropriate specificity (not too vague or overly detailed).

  • Reason for selection is mentioned.

What you do next: Accept or Ask the agent to add any specific requirements.

 

Step 3: Upload your dataset

Step 3: Upload your dataset

Actions:

  1. In the chat, click the attachment icon (📎).

  2. Select your images_with_queries.json file.

  3. Wait for the agent to process it.

image (18).png

 

What the agent responds:

  • Confirms file upload.

  • Shows detected fields and structure.

image (19).png

 

What you check: ✅

  • Image links are detected correctly.

  • Metadata fields are recognized.

  • Both suggested queries are detected.

  • All fields are present.

What you do next: Accept or ask the agent to fix.

Step 4: Designing datapoint structure

Step 4: Designing datapoint structure

Actions:

  1. Review agent's proposed structure.

  2. Define input entities (what experts see):

    • image_link (image URL);

    • query1 (first suggested query);

    • query2 (second suggested query).

  3. Define output entities (what experts create):

    • selected_query (the chosen query);

    • selection_reason (brief explanation).

  4. Link output entities to input entities.

 

What the agent proposes:

  • Input/output entity structure based on your dataset.

image (20).png

 

What you check:

All input entities exist (both queries shown as inputs).

image (21).png
  • All output entities exist (selected query AND selection reasoning).

Entity types correct (image link as IMAGE type).

image (22).png

Extracted data correct (for input entities, e.g. check if the “image link” is extracted correctly).

image (23).png

 

What you do next: Accept or ask the agent to fix.

Step 5: Defining quality criteria

Step 5: Defining quality criteria

Actions:

  1. Review agent's proposed quality criteria.

  2. Verify criteria cover:

    • Selection accuracy according to selection criteria.

    • Reasoning quality.

    • Dataset-aware selection.

What the agent proposes:

  • Quality Requirements with pass/fail conditions.

  • Criteria aligned with your selection rules.

    image (24).png

What you check: ✅

  • Quality Requirements cover selection accuracy according to the selecting criteria, reasoning quality.

Verify that quality element rules are realistic and achievable.

image (25).png

Generated quality element clearly show pass/fail conditions.

image (26).png

Criteria align with your selection rules.

image (27).png

 

What you do next: Accept or ask the agent to fix.

image (28).png

 

What the agent confirms:

✅ Quality Requirements configured.

These requirements will be used by LLM QA to automatically check expert work.
Experts must meet these standards for their annotations to pass.

Ready to proceed to Expert Guidelines?

What you do next: Accept or ask the agent to fix.

 

Step 6: Building annotation interface

Step 6: Building annotation interface

Actions:

  • Create user friendly UI for experts.

What the agent proposes

  • Project UI for experts.

    image (29).png

What you check: ✅

  • The annotation flow is logical: the image appears first, search queries follow, comparison criteria are visible, selection options are named consistently, and expectations for the reasoning field are clearly explained.

  • Search queries are placed side by side.

What you do next: Accept or ask the agent to fix the UI

Fixing example

image (31).png

  • Review the updates, then Accept;

  • or continue refining until satisfied.

    image (30).png

Step 7: Writing annotator guidelines

Step 7: Writing annotator guidelines

Actions:

  1. Write clear instructions for experts:

    • How to evaluate queries.

    • Selection criteria.

    • How to write selection reason.

  2. Add examples:

    • Good selection with explanation.

    • Bad selection with explanation.

  3. Include edge cases and how to handle them

What the agent generates:

image (32).png

 

What you check: ✅

Guidelines clearly explain the selection task workflow and match the UI flow.

image (33).png

Selection rules are detailed and match your requirements.

image (34).png

Examples show the evaluation and selection process.

image (2).png

Edge cases are covered.

image (3).png

Quality rubrics are provided.

image (4).png

 

What you do next: Accept or ask the agent to fix the guideline

Click “Proceed”.

Step 8: Self-check (test your configuration)
Step 9: Set pricing and launch

Step 9: Set pricing and launch

Actions:

Agent recommends you several options with suitable experts. Choose your audience.

image (13).png

Once the audience chosen, proceed to confirm cost estimates by clicking “Save”.

image (14).png

Select number of tasks, estimate task completion time, set the price for completing one task.

image (15).png

  • Click "Launch" to start labeling.

Step 10: Monitor results

Step 10: Monitor results

What you see:

  • Tasks sent for annotation

  • Status for each task:

    • Finding Expert - the task is available on the tasks marketplace

    • In Expert Work - an expert is working on the task

    • Quality Check - LLM QA is checking the task

    • Ready for Review - the task is waiting for your review

    image (16).png

Actions:

  1. Wait (experts are working on your tasks - evaluation and selection takes time). It can takes 30-60 min.

  2. Track item statuses in dashboard:

    • Finding Expert → In Expert Work → Quality Check → Ready for Review → Completed.

  3. Review results with status "Ready for Review".

  4. Check LLM QA evaluations.

  5. Monitor progress metrics.

What you check: ✅

  • Results are coming in.

  • Each result has a selected query and reason.

  • Quality checks are passing.

Step 11: Do user quality calibration

Step 11: Do user quality calibration

Actions:

  1. Review results in "Ready for review" status.

Click on a completed item to review it.

image (17).png

  1. Evaluate each Quality Criterion (Pass/Fail).

  • LLM QA evaluation on quality elements (the naming may vary depending on how you configured):

    • Selection Accuracy on Selected Query.

    • Reasoning Quality on Selection Reasoning.

image (18).png

  1. Write reasoning for each evaluation.

  2. Compare your results with LLM QA assessments.

    Check & Provide Quality Feedback: ✅

    • Check if a search query is selected (radio group is toggled)

    • Check if the selection reason is provided and explains the choice

    • Review LLM comments on each quality elements (Selection Accuracy on Selected Query and Reasoning Quality on Selection Reasoning)

      • In early stages, review as many tasks as you want

      • The review results will be considered in the next project iteration

      • You'll be asked to leave comments if you disagree with the LLM QA verdict

        image (19).png image (20).png

  3. If satisfied Download results as JSON

    image (21).png

  4. If you want to improve quality, complete review and create new project version, Agent will help you to change the settings based on your quality feedback.

    image (22).png

    • What the agent proposes:

image (23).png

  1. The agent will guide you through the next iteration of your project configuration.

 

Use Case: search query generation

Experts generate two distinct search queries for each image, focusing on natural and implicit user intent rather than literal visual details. Each query should be grammatically correct, context-aware, and tailored for searching within a curated dataset that contains diverse but similar image types without exact duplicates. The queries must avoid being overly vague or excessively specific, striking a balance that helps identify the image’s purpose or use case.

⏱️ 30-60 min | 💰 ~$3-4 per task | 👥 3400+ experts available

 

Step 1: Prepare your dataset

Step 1: Prepare your dataset

Actions:

  1. Create a JSON file with your images links.

  2. Save it as image_links.json.

  3. Ensure image URLs are publicly accessible.

Or use example files:

  1. Example file

Your dataset structure (image_links.json)

[
  {
    "id": "KmQpXvLa",
    "link": "<https://prodtlkcsafiles.blob.core.windows.net/self-service-
    images-comparison/pair-0001-img-1.png>"
  },
  {
    "id": "RtYnBwEh",
    "link": "<https://prodtlkcsafiles.blob.core.windows.net/self-service-
    images-comparison/pair-0002-img-1.png>"
  },
]

Check: Valid JSON, all fields filled, image URLs accessible.

Step 2: Create project and describe your task

Step 2: Create project and describe your task

Actions:

  1. Log into the platform.

  2. Click "Create new project" to create project within a project space.

image (25).png

  1. Describe your task to the agent What to type:

I need experts to generate two distinct search queries for each image that would 
realistically help a user find that specific image within a curated dataset. The dataset 
contains diverse images of similar types and topics (such as dashboards, data 
visualizations, and analytics interfaces), but no exact duplicates. Each query should 
be crafted as if a real user were searching for the image, balancing specificity and 
naturalness.

Query requirements:
- Natural and grammatically correct – phrased like a realistic user request, 
not a list of keywords
- Implicit rather than literal – describe the image’s purpose, concept, or use-case 
rather than listing visible UI elements or colors
- Dataset-aware specificity – include enough detail to distinguish from similar images 
in the dataset, but avoid excessive or made-up details
- Not too vague – generic terms like “dashboard” or “analytics” alone are insufficient
- Not overly specific – avoid describing exact layouts, specific percentages, 
or granular visual components

For each image, experts should provide two queries that approach the image from 
different angles (e.g., business context vs. user persona), ensuring both would plausibly 
retrieve the target image from the dataset.

Or use description from: Project description file.

 

On the right side, you'll chat with the AI Assistant. After each step, you'll be invited to Accept or modify the generated results shown in the main window on the left. Once you Accept, you'll see the agent's report on the completed workflow. To view the next step's results, you must also accept the configured content.

image (26).png

 

What you check: ✅

  • The project goal mentions GENERATION of queries.

  • The annotation task emphasizes query generation.

  • Inputs and outputs are mentioned.

  • Deliverables.

  • Generated queries criteria are listed.

What You Do Next

Accept or Ask the agent to add any specific requirements.

  • In this case, since we just need two generated queries, the agent was informed accordingly.

image (27).png

image (28).png

  • Accept when satisfied.

image (29).png

Step 3: Upload your dataset

Step 3: Upload your dataset

Actions:

  1. In the chat, click the attachment icon (📎).

  2. Select your image_links.json file.

  3. Wait for the agent to process.

image (2).png

 

What the agent responds:

  • Confirms file upload

  • Shows detected fields and structure

image (3).png

 

What you check: ✅

  • Image links are detected correctly.

  • Number of items is correct.

 

What you do next: Accept or ask the agent to fix.

Step 4: Designing datapoint structure

Step 4: Designing datapoint structure

Actions:

  1. Review agent's proposed structure.

  2. Define input entities (what experts see).

  3. Define output entities (what experts create).

  4. Link output entities to input entities.

What the agent proposes:

  • Input/output entity structure based on your dataset

image (4).png

 

What you check: ✅

Check if all input and output entities exist, e.g. image link and two queries.

image (5).png

Check input and output entities types, e.g., “image link” is configured as IMAGE entity type or you can ask why some entities have a certain type.

image (6).png

image (7).png

Check extracted data for input entities, e.g. check if the “image link” is extracted correctly.

image (8).png

What you do next: Accept or ask the agent to fix.

Step 5: Defining quality criteria

Step 5: Defining quality criteria

Actions:

  1. Review agent's proposed quality criteria.

  2. Verify criteria cover:

    • Implicit style, naturalness, specificity balance, and word count of generated queries.

    • Distinctness between the two queries.

What the agent proposes:

  • Quality Requirements with pass/fail conditions.

  • Criteria aligned with your selection rules.

image (9).png

 

What you check: ✅

  • Quality Requirements address the criteria for good queries.

Query distinctness is addressed.

image (10).png

  • Generated quality element clearly show pass/fail conditions.

 

What the agent confirms:

✅ Quality Requirements configured.

These requirements will be used by LLM QA to automatically check expert work.
Experts must meet these standards for their annotations to pass.

Ready to proceed to Expert Guidelines?

What you do next: Accept or ask the agent to fix.

Step 6: Building annotation interface

Step 6: Building annotation interface

Actions:

  • Create user friendly UI for experts.

What the agent generates:

  • Project UI for experts.

image (11).png

 

What you check: ✅

  • The workflow is logical: the image appears next to expected search queries, the criteria are visible.

  • The UI is user-friendly and easy to navigate.

What you do next: Accept or ask the agent to fix the UI.

Fixing Example

  • For example, are all criteria for queries displayed for visibility.

    image (12).png

    • Review options and choose if you agree.

    • Review the updates, then accept or continue refining until satisfied.

    image (13).png

Step 7: Writing annotator guidelines

Step 7: Writing annotator guidelines

Actions:

  1. Write clear instructions for experts.

  2. Add examples.

  3. Include edge cases and how to handle them.

 

What the agent generates:

image (14).png

 

What You Check: ✅

Guidelines clearly explain the selection task workflow and match the UI flow.

image (15).png

  • Quality criteria explained.

  • Examples are provided.

  • The guideline sections are logical and easy to understand.

 

What you do next: Accept or ask the agent to fix the guideline. Click “Proceed”.

Step 8: Self-check (test your configuration)

Step 8: Self-check (test your configuration)

Actions:

  1. Complete at least 1-2 tasks from 10 sample tasks yourself.

  2. Test the interface, guidelines, and quality criteria.

  3. Review LLM QA evaluations.

 

What You check:

  • Interface usability.

  • Task completeness.

  • LLM QA responses.

  • Guideline clarity.

 

What you do:

  1. Click "Proceed" to navigate to the Self-check step.

  2. The system shows you 10 sample tasks from your dataset.

  3. Expand the task to review the UI.

  4. Check that all inputs are extracted correctly in the provided samples: images display properly, fields for queries are shown, etc.

  5. Complete the first task yourself:

    • View the image.

    • Write queries.

    • Submit your answer by clicking on the ”Submit to LLM QA”.

    • LLM QA evaluation will take 30-60 sec.

  6. Verify that the LLM evaluation and comments align with the pass or fail criteria. If not, go to Project Configuration step and improve project settings.

  7. You can download completed test samples with status “Ready for Review” by selecting them.

  8. You can test as many cases as you want.

    1. Test very short reasoning that you can limit before LLM QA.

    2. Repeating the query in the reasoning.

What You Do Next: Complete 2-3 more test tasks with different image types to ensure everything works correctly.

  • If not, go to the Project Configuration step and improve the settings.

  • If yes, go to the Contract section by clicking “Proceed”

Step 9: Set pricing and launch

Step 9: Set pricing and launch

Actions:

Agent recommends you several options with suitable experts. Choose your audience.

image (16).png

Once the audience chosen, proceed to confirm cost estimates by clicking “Save”.

image (17).png

Select number of tasks, estimate task completion time, set the price for completing one task.

image (18).png

  • Click "Launch" to start labeling

Step 10: Monitor results

Step 10: Monitor results

What you see:

  • Tasks sent for annotation.

  • Status for each task:.

    • Finding Expert - the task is available on the tasks marketplace.

    • In Expert Work - an expert is working on the task.

    • Quality Check - LLM QA is checking the task.

    • Ready for Review - the task is waiting for your review.

image (19).png

 

Actions:

  1. Wait (experts are working on your tasks - evaluation and selection takes time). It can takes 30-60 min.

  2. Track item statuses in dashboard:

    • Finding Expert → In Expert Work → Quality Check → Ready for Review → Completed

  3. Review results with status "Ready for Review".

  4. Check LLM QA evaluations.

  5. Monitor progress metrics.

 

What you check: ✅

  • Results are coming in.

  • Each result has a selected query and reason.

  • Quality checks are passing.

Step 11: Do user quality calibration

Step 11: Do user quality calibration

What you do:

  1. Review results in "Ready for Review" status.

  2. Evaluate each Quality Criterion (Pass/Fail).

    • LLM QA evaluation on quality elements (the naming may vary depending on how you configured).

  3. Write reasoning for each evaluation.

  4. Compare your results with LLM QA assessments.

Check & provide quality feedback: ✅

  • Check if a search query is selected (radio group is toggled).

  • Check if the selection reason is provided and explains the choice.

  • Review LLM comments on each quality elements.

    • In early stages, review as many tasks as you want.

    • The review results will be considered in the next project iteration.

    • You'll be asked to leave comments if you disagree with the LLM QA verdict.

  1. If satisfied Download results as JSON.

  2. If you want to improve quality, complete review and create new project version, Agent will help you to change the settings based on your quality feedback.

What the agent proposes:

image (20).png

  1. The agent will guide you through the next iteration of your project configuration

 

Use Case: Image classification

This project classifies images into predefined categories based on their visual intent and communicative purpose. Inputs include image URLs and IDs. The categories include: Graphs and Charts (data visualizations), Diagrams (technical schematics, categorization structures, timelines, product instructions), Data and Process Flows (workflows, pipelines, decision pathways), Infographics (rich visual summaries and dashboards), and Multi-Figure Images (composite panels from reports or studies). Images that do not fit any category are labeled as Uncategorized. The output is a JSON file pairing each image’s link and ID with its assigned category, enabling consistent organization and downstream analysis of visual content.

⏱️ 30-60 min | 💰 ~$3-4 per task | 👥 3400+ experts available

 

Step 1: Prepare your dataset

Step 1: Prepare your dataset

What you do:

  1. Create a JSON file with your images links.

  2. Save it as image_links.json.

  3. Ensure image URLs are publicly accessible.

Or use example files:

  1. Example file

Your dataset structure (image_links.json).

[
  {
    "id": "KmQpXvLa",
    "link": "<https://prodtlkcsafiles.blob.core.windows.net/self-service-images-
    comparison/pair-0001-img-1.png>"
  },
  {
    "id": "RtYnBwEh",
    "link": "<https://prodtlkcsafiles.blob.core.windows.net/self-service-images-
    comparison/pair-0002-img-1.png>"
  },
]

Step 2: Create project and describe your task

Step 2: Create project and describe your task

Actions:

  1. Log into the platform.

  2. Create Project Space (e.g., "My First Projects").

  3. Click "Create New Project" to Create Project within a Project Space.

image (21).png

  1. Describe your task to the agent What you type:

I need experts to classify a list of images into predefined categories based on their 
visual content and purpose. The input consists of image links and image IDs. Each image 
should be analyzed and assigned to exactly one of the following categories:

Classification Categories:

- Graphs and Charts – Visual representations of data highlighting trends, patterns, and 
relationships through graphical elements like lines, bars, and pie charts.

- Diagrams – Including the following subtypes:
	* Technical diagrams: Network architecture, system components, and detailed 
	schematic representations of complex systems, equipment, or processes, often used 
	in engineering, manufacturing, and technical fields
	* Categorization diagrams: Illustrations that organize and classify information 
	into distinct groups or hierarchies, such as concept maps, taxonomies, or 
	organizational 
	structures
	* Hierarchical and timeline diagrams: Visuals showing relationships, structures, or 
	chronological order, including organizational charts, family trees, and project 
	timelines
	* Product diagrams/images from user manuals: Illustrations or photos depicting 
	components, features, or usage of a product, typically found in user guides, 
	instruction manuals, or product documentation

- Data and Process Flow – Diagrams portraying the movement and transformation of data 
or the sequence of steps in a workflow or process, including decision trees; commonly 
used in software engineering, business analysis, and process improvement.

- Infographics – Visually engaging graphics combining data, information, and design 
elements to communicate concepts, trends, or detailed information in a digestible format, 
including data visualizations and dashboards.

- Multi-Figure Images – Composed of multiple related figures, such as side-by-side 
panels or labeled figures from documentation or reports.

- Uncategorized – Any image not matching the above categories.

Classification requirements:

Mutually exclusive – each image should be assigned to only one category, even if 
it contains elements from multiple categories; choose the most dominant or primary purpose

Content-driven – base the classification on the image’s visual content and intended 
purpose, not on file names or metadata

Subtype-aware – for images classified as “Diagrams,” identify the most appropriate 
subtype based on the image’s structure and use-case

Consistent application – apply the same criteria uniformly across all images to ensure 
classification consistency

When in doubt – if an image does not clearly fit any category, label it as
“Uncategorized” rather than forcing a fit

Output:
For each image, experts should provide a JSON entry containing:
Image link
Image ID
Assigned category (and subtype if applicable for Diagrams

Or use description from: Project description file.

 

On the right side, you'll chat with the AI Assistant. After each step, you'll be invited to Accept or modify the generated results shown in the main window on the left. Once you Accept, you'll see the agent's report on the completed workflow. To view the next step's results, you must also accept the configured content.

image (22).png

 

What you check: ✅

  • The project goal mentions image classification.

  • Inputs and outputs are mentioned.

  • Image categories are listed.

  • Deliverables.

  • Quality expectations.

What you do next: Accept or ask the agent to add any specific requirements.

Step 3: Upload your dataset

Step 3: Upload your dataset

Actions:

  1. In the chat, click the attachment icon (📎).

  2. Select your image_links.json file.

  3. Wait for the agent to process.

image (23).png

 

What the agent responds:

  • Confirms file upload.

  • Shows detected fields and structure.

image (24).png

 

What you check: ✅

  • Image links are detected correctly.

  • Number of items is correct.

What you do next: Accept or ask the agent to fix.

Step 4: Designing datapoint structure

Step 4: Designing datapoint structure

Actions:

  1. Review agent's proposed structure.

  2. Define input entities (what experts see).

  3. Define output entities (what experts create).

  4. Link output entities to input entities.

 

What the agent proposes:

  • Input/output entity structure based on your dataset.

image (25).png

 

What you check: ✅

Check if all input and output entities exist, e.g. image link and two queries.

image (26).png

Check input and output entities types, e.g., “image link” is configured as IMAGE entity type or you can ask why some entities have a certain type.

image (27).png

Check extracted data for input entities, e.g. check if the “image link” is extracted correctly.

image (28).png

 

What you do next: Accept or ask the agent to fix.

Step 5: Defining quality criteria

Step 5: Defining quality criteria

Actions:

  1. Review agent's proposed quality criteria.

  2. Verify criteria cover:

    • Classification accuracy.

    • Reasoning quality.

 

What the agent proposes:

  • Quality Requirements with pass/fail conditions.

  • Criteria aligned with your selection rules.

image (29).png

 

What you check: ✅

  • Check if proposed quality criteria are not redundant.

Request the agent to edit criteria.

image (2).png

  • Quality Requirements address classification accuracy.

Generated quality element clearly show pass/fail conditions.

image (3).png

What you do next: Accept or ask the agent to fix.

 

What the agent confirms:

✅ Quality Requirements configured.

These requirements will be used by LLM QA to automatically check expert work.
Experts must meet these standards for their annotations to pass.

Ready to proceed to Expert Guidelines?
Step 6: Building annotation interface

Step 6: Building annotation interface

Actions:

  • Create user friendly UI for experts.

What the agent proposes

  • Project UI for experts.

    image (4).png

What you check: ✅

  • The workflow is logical: the image appears next to the categories options, explanation text field is available, category definition are available.

  • The UI is user-friendly and easy to navigate.

What you do next: Accept or ask the agent to fix the UI.

Step 7: Writing annotator guidelines

Step 7: Writing annotator guidelines

Actions:

  1. Write clear instructions for experts.

  2. Add examples.

  3. Include edge cases and how to handle them.

 

What the agent generates:

image (5).png

What you check: ✅

Guidelines clearly explain the selection task workflow and match the UI flow

image (6).png

  • Category definitions are provided.

  • Quality criteria explained.

  • Examples are provided.

  • The guideline sections are logical and easy to understand.

 

What you do next: Accept or ask the agent to fix the guideline. Click “Proceed”.

Step 8: Self-check (test your configuration)

Step 8: Self-check (test your configuration)

Actions:

  1. Complete at least 1-2 tasks from 10 sample tasks yourself.

  2. Test the interface, guidelines, and quality criteria.

  3. Review LLM QA evaluations.

What you check:

  • Interface usability.

  • Task completeness.

  • LLM QA responses.

  • Guideline clarity.

What you do:

  1. Click "Proceed" to navigate to the Self-Check step.

  2. The system shows you 10 sample tasks from your dataset.

  3. Expand the task to review the UI.

  4. Check that all inputs are extracted correctly in the provided samples: images display properly, fields for queries are shown, etc.

  5. Complete the first task yourself based on the UI:

    • View the image.

    • Write queries.

    • Submit your answer by clicking on the ”Submit to LLM QA”.

    • LLM QA evaluation will take 30-60 sec.

  6. Verify that the LLM evaluation and comments align with the pass or fail criteria. If not, go to Project Configuration step and improve project settings.

  7. You can download completed test samples with status “Ready for Review” by selecting them.

What you do next: Complete 2-3 more test tasks with different image types to ensure everything works correctly.

  • If not, go to the Project Configuration step and improve the settings;

  • If yes, go to the Contract section by clicking “Proceed”.

Step 9: Set pricing and launch

Step 9: Set pricing and launch

Actions:

Agent recommends you several options with suitable experts. Choose your audience.

image (7).png

Once the audience chosen, proceed to confirm cost estimates by clicking “Save”.

image (8).png

Select number of tasks, estimate task completion time, set the price for completing one task.

image (9).png
  • Click "Launch" to start labeling.

Step 10: Monitor results

Step 10: Monitor results

What you see:

  • Tasks sent for annotation.

  • Status for each task:

    • Finding Expert - the task is available on the tasks marketplace.

    • In Expert Work - an expert is working on the task.

    • Quality Check - LLM QA is checking the task.

    • Ready for Review - the task is waiting for your review.

Actions:

  1. Wait (experts are working on your tasks - evaluation and selection takes time). It can takes 30-60 min.

  2. Track item statuses in dashboard:

    • Finding Expert → In Expert Work → Quality Check → Ready for Review → Completed

  3. Review results with status "Ready for Review".

  4. Check LLM QA evaluations.

  5. Monitor progress metrics.

What you check: ✅

  • Results are coming in.

  • Each result has a selected query and reason.

  • Quality checks are passing.

Step 11: Do user quality calibration

Step 11: Do user quality calibration

Actions:

  1. Review results in "Ready for Review" status.

    • Click on a completed item to review it.

image (10).png

  1. Evaluate each Quality Criterion (Pass/Fail).

    • LLM QA evaluation on quality elements (the naming may vary depending on how you configured).

  2. Write reasoning for each evaluation.

  3. Compare your results with LLM QA assessments.

Check & provide quality feedback: ✅

Review LLM comments on each quality elements (Selection Accuracy on Selected Query and Reasoning Quality on Selection Reasoning).

  • In early stages, review as many tasks as you want.

  • The review results will be considered in the next project iteration.

  • You'll be asked to leave comments if you disagree with the LLM QA verdict.

image (11).png
image (12).png
  1. If satisfied Download results as JSON

  2. If you want to improve quality, complete review and create new project version, Agent will help you to change the settings based on your quality feedback

image (13).png
  1. The agent will guide you through the next iteration of your project configuration.

 

Use Case: Images comparison

Experts evaluate two suggested images and select the one that best matches specific criteria.

⏱️ 30-60 min | 💰 ~$1 per task| 👥 3400+ experts available

 

Step 1: Prepare your dataset

Step 1: Prepare your dataset

Actions:

  1. Create a JSON file with images, metadata, and two suggested queries per image.

  2. Save as images_to_compare.json.

  3. Ensure image URLs are publicly accessible.

Or use example files:

  • Example file

  • Your dataset structure (images_to_compare.json).

    [
      {
        "pair_id": "pair-0001",
        "image_1": "<https://prodtlkcsafiles.blob.core.windows.net/self-service-images-
        comparison/pair-0001-img-1.png>",
        "image_2": "<https://prodtlkcsafiles.blob.core.windows.net/self-service-images-
        comparison/pair-0001-img-2.png>"
      },
      {
        "pair_id": "pair-0002",
        "image_1": "<https://prodtlkcsafiles.blob.core.windows.net/self-service-images-
        comparison/pair-0002-img-1.png>",
        "image_2": "<https://prodtlkcsafiles.blob.core.windows.net/self-service-images-
        comparison/pair-0002-img-2.png>"
      }
    ]

Check: Valid JSON, all fields filled, image URLs accessible.

Step 2: Create project and describe your task

Step 2: Create project and describe your task

Actions:

  1. Log into the platform.

  2. Create Project Space (e.g., "My First Projects").

  3. Click "Create New Project" to Create Project within a Project Space.

    image (14).png

  4. Describe your task to the agent.

 

What to type:

I need experts who can assess and compare two given images based on standardized 
evaluation metrics. The goal is to measure clarity, design quality, completeness, and 
overall communication effectiveness. The image receiving the highest total score will be 
declared the winner.

Evaluation Metrics (0–5 points each)

1. Title & Labelling
Measures how clear and descriptive the image's title and labels are.
- 0-2: Missing title and/or labels
- 3-4: Present but unclear/incomplete
- 5: Clear, descriptive title and well-labelled components

2. Grammar & Language
Assesses spelling, grammar, and professionalism of text in the image.
- 0: Multiple grammar/spelling errors
- 3-4: Minor errors not hindering understanding
- 5: Grammatically correct with professional wording

3. Clarity of Process Depiction
Evaluates how clearly the image presents steps or concepts in logical order.
- 0–2: Steps unclear or missing
- 3–4: Structure mostly clear with minor gaps
- 5: Fully clear, chronologically correct, and visually guided

4. Standalone Understanding
Judges whether the image can be understood without external explanations.
- 0–2: Needs heavy external context
- 3–4: Understandable with some background knowledge
- 5: Understandable to a layperson

5. Visual Design & Readability
Considers layout, color contrast, and visual appeal.
- 0-2: Poor design and color use
- 3-4: Readable but slightly messy or inconsistent
- 5: Clean, appealing, and visually consistent
6. Completeness & Key Content Coverage
Looks at whether all major elements are included.
- 0–2: Many missing elements
- 3–4: Most concepts covered but incomplete
- 5: All major items included accurately

Scoring System
- Each metric scores 0–5 points
- Maximum possible score per image: 30 points
- The image with the highest score wins
- Tiebreaker: Overall clarity and/or human preference

The experts evaluate both images against all metrics and record the scores in a 
comparison table. Calculate totals to determine the winner, then assign a descriptive 
title to the winning image.

Or use description from: Project Description File.

 

On the right side, you'll chat with the AI Assistant. After each step, you'll be invited to Accept or modify the generated results shown in the main window on the left. Once you Accept, you'll see the agent's report on the completed workflow. To view the next step's results, you must also accept the configured content.

image (15).png

 

What you check: ✅

  • The project goal mentions SELECTING the best image.

  • The annotation task emphasizes evaluation and selection.

  • Evaluation criteria are mentioned.

  • Scoring system is clearly explained.

What you do next: Accept or Ask the agent to add any specific requirements.

Step 3: Upload your dataset

Step 3: Upload your dataset

Actions:

  1. In the chat, click the attachment icon (📎).

  2. Select your images_to_compare.json file.

    image (16).png

  3. Wait for the agent to process it.

 

What the agent responds:

  • Confirms file upload.

  • Shows detected fields and structure.

image (17).png

 

What you check: ✅

  • Images links are detected correctly.

 

What you do next: Accept or ask the agent to fix.

Step 4: Designing datapoint structure

Step 4: Designing datapoint structure

Actions:

  1. Review agent's proposed structure.

  2. Define input entities (what experts see).

  3. Define output entities (what experts create).

 

What the agent proposes:

  • Input/output entity structure based on your dataset.

image (18).png

 

What you check:

All input entities exist (both queries shown as inputs).

image (19).png

All output entities exist (selected query AND selection reasoning).

image (20).png

Entity types correct (image link as IMAGE type).

image (21).png

Extracted data correct (for input entities, e.g. check if the “image link” is extracted correctly).

image (22).png

 

What you do next: Accept or ask the agent to fix.

Step 5: Defining quality criteria

Step 5: Defining quality criteria

Actions:

  1. Review agent's proposed quality criteria.

  2. Verify criteria cover:

    • Evaluation metrics.

    • Reasoning quality.

 

What the agent proposes:

  • Quality requirements with pass/fail conditions.

  • Criteria aligned with your selection rules.

image (23).png

 

What you check: ✅

  • Quality Requirements cover all metrics.

Request the agent to edit criteria.

image (24).png

Verify that quality element rules are realistic and achievable.

image (25).png

Generated quality element clearly shows pass/fail conditions.

image (26).png

Criteria align with your selection rules.

image (27).png

What you do next: Accept or ask the agent to fix.

 

What the agent confirms:

🔎 **Review the criteria** to verify pass/fail thresholds align with your evaluation 
standards.

Approve to continue → **Building annotation interface** — or suggest changes

 

What you do next: Accept or ask the agent to fix.

Step 5: Defining quality criteria

Step 5: Defining quality criteria

Actions:

  1. Review agent's proposed quality criteria.

  2. Verify criteria cover:

    • Evaluation metrics.

    • Reasoning quality.

 

What the agent proposes:

  • Quality requirements with pass/fail conditions.

  • Criteria aligned with your selection rules.

image (23).png

 

What you check: ✅

  • Quality Requirements cover all metrics.

Request the agent to edit criteria.

image (24).png

Verify that quality element rules are realistic and achievable.

image (25).png

Generated quality element clearly shows pass/fail conditions.

image (26).png

Criteria align with your selection rules.

image (27).png

What you do next: Accept or ask the agent to fix.

 

What the agent confirms:

🔎 **Review the criteria** to verify pass/fail thresholds align with your evaluation 
standards.

Approve to continue → **Building annotation interface** — or suggest changes

 

What you do next: Accept or ask the agent to fix.

Step 7: Writing annotator guidelines

Step 7: Writing annotator guidelines

Actions:

  1. Write clear instructions for experts:

    • Workflow description

    • Evaluation metrics and scoring rubrics

    • Reasoning expectations

  2. Add examples:

    • Pass example

    • Fail example

 

What the agent generates:

image (30).png

 

What you check: ✅

Guidelines clearly explain the selection task workflow and match the UI flow.

image (31).png

Evaluation metrics and scoring rubrics are explained in detail.

image (32).png

Examples show the evaluation and selection process.

image (33).png

Quality rubrics are provided.

image (34).png

 

What you do next: Accept or ask the agent to fix the guideline. Click “Proceed”.

Step 8: Self-check (test your configuration)

Step 8: Self-check (test your configuration)

Actions:

  1. Complete at least 1-2 tasks from 10 sample tasks yourself.

  2. Test the interface, guidelines, and quality criteria.

  3. Review LLM QA evaluations.

 

What you check:

  • Interface usability.

  • Task completeness.

  • LLM QA responses.

  • Guideline clarity.

 

What you do:

  1. Click "Proceed" to navigate to the Self-Check step.

  2. The system shows you 10 sample tasks from your dataset.

  3. Expand the task to review the UI.

  4. Check that all inputs are extracted correctly in the provided samples: images display properly, evaluation metric fields are available, etc.

  5. Complete the first task yourself:

    • Submit your answer by clicking on the ”Submit to LLM QA”.

    • LLM QA evaluation will take 30-60 sec.

  6. Verify that the LLM evaluation and comments align with the pass or fail criteria. If not, go to Project Configuration step and improve project settings.

  7. You can download completed test samples with status “Ready for Review” by selecting them.

  8. You can test as many cases as you want.

 

What you do next: Complete 2-3 more test tasks with different image types to ensure everything works correctly.

  • If not, go to the Project Configuration step and improve the settings.

  • If yes, go to the Contract section by clicking “Proceed”.

Step 9: Set pricing and launch

Step 9: Set pricing and launch

Actions:

Agent recommends you several options with suitable experts. Choose your audience.

image (2).png

Once the audience chosen, proceed to confirm cost estimates by clicking “Save”.

Select number of tasks, estimate task completion time, set the price for completing one task.

image (3).png

Click "Launch" to start labeling.

Step 10: Monitor results

Step 10: Monitor results

What you see:

image (4).png

  • Tasks sent for annotation.

  • Status for each task:

    • Finding Expert - the task is available on the tasks marketplace.

    • In Expert Work - an expert is working on the task.

    • Quality Check - LLM QA is checking the task.

    • Ready for Review - the task is waiting for your review.

 

Actions:

  1. Wait (experts are working on your tasks - evaluation and selection takes time). It can takes 30-60 min.

  2. Track item statuses in dashboard:

    • Finding Expert → In Expert Work → Quality Check → Ready for Review → Completed

  3. Review results with status "Ready for Review".

  4. Check LLM QA evaluations.

  5. Monitor progress metrics.

 

What you check: ✅

  • Results are coming in.

  • Each result has a selected image and reason.

  • Quality checks are passing.

Step 11: Do user quality calibration

Step 11: Do user quality calibration

Actions:

  1. Review results in "Ready for Review" status.

    • Click on a completed item to review it.

      image (5).png

  2. Evaluate each Quality Criterion (Pass/Fail).

    • LLM QA evaluation on quality elements (the naming may vary depending on how you configured).

  3. Write reasoning for each evaluation.

  4. Compare your results with LLM QA assessments.

 

Check & Provide Quality Feedback: ✅

  • Check if a search query is selected (radio group is toggled).

  • Check if the selection reason is provided and explains the choice.

Review LLM comments on each quality elements (Selection Accuracy on Selected Query and Reasoning Quality on Selection Reasoning).

  • In early stages, review as many tasks as you want.

  • The review results will be considered in the next project iteration.

  • You'll be asked to leave comments if you disagree with the LLM QA verdict.

  1. If satisfied Download results as JSON.

  2. If you want to improve quality, complete review and create new project version, Agent will help you to change the settings based on your quality feedback.

    image (6).png

  3. The agent will guide you through the next iteration of your project configuration.

 

Use Case: Image-based question and answer generation

Experts analyze provided images and create one question-answer pair per image. Each question is categorized by type: Lookup (finding visible information), Summarization (describing content), or Analytical (reasoning/calculation based on image data). The goal is to generate diverse, high-quality Q&A pairs that test different levels of image understanding.

⏱️ 30-60 min | 💰 ~$3 per task| 👥 3400+ experts available

 

Step 1: Prepare your dataset

Step 1: Prepare your dataset

Actions:

  1. Create a JSON file with images, metadata, and two suggested queries per image.

  2. Save as image_links.json.

  3. Ensure image URLs are publicly accessible.

Or use example files:

Your Dataset Structure (image_links.json)

[
  {
    "id": "KmQpXvLa",
    "link": "<https://prodtlkcsafiles.blob.core.windows.net/self-service-images
    -comparison/pair-0001-img-1.png>"
  },
  {
    "id": "RtYnBwEh",
    "link": "<https://prodtlkcsafiles.blob.core.windows.net/self-service-images
    -comparison/pair-0002-img-1.png>"
  }
]

Check: Valid JSON, all fields filled, image URLs accessible

Step 2: Create project and describe your task

Step 2: Create project and describe your task

Actions:

  1. Log into the platform.

  2. Create Project Space (e.g., "My First Projects").

  3. Click "Create New Project" to Create Project within a Project Space.

image (7).png

  1. Describe your task to the agent What you type:

I need experts to create high-quality question-answer pairs based on provided images. 
For each image link, experts should generate one question and its corresponding answer.
For each image, I need experts to provide:
1. Question Type (select one):
- Lookup – A question that requires finding specific, directly visible information in the 
image
- Summarization – A question that requires describing, interpreting, or summarizing the 
image content
- Analytical/Reasoning – A question that requires calculation, inference, or logical 
reasoning based on image data
2. Question – A clear, well-formed question about the image
3. Answer – The accurate answer to the question

Or use description from: Project description file.

On the right side, you'll chat with the AI Assistant. After each step, you'll be invited to Accept or modify the generated results shown in the main window on the left. Once you Accept, you'll see the agent's report on the completed workflow. To view the next step's results, you must also accept the configured content.

image (8).png

 

What you check: ✅

  • The project goal mentions question and answer generation.

  • Inputs and outputs are mentioned.

  • Question categories are listed.

  • Quality expectations are provided.

What you do next: Accept or Ask the agent to add any specific requirements.

Step 3: Upload your dataset

Step 3: Upload your dataset

Actions:

  1. In the chat, click the attachment icon (📎).

  2. Select your image_links.json file.

  3. Wait for the agent to process.

What the agent responds:

  • Confirms file upload.

  • Shows detected fields and structure.

image (9).png

 

What you check: ✅

  • Image links are detected correctly

  • Number of items is correct

What You Do Next: Accept or ask the agent to fix

Step 4: Designing datapoint structure

Step 4: Designing datapoint structure

Actions:

  1. Review agent's proposed structure.

  2. Define input entities (what experts see).

  3. Define output entities (what experts create).

  4. Link output entities to input entities.

 

What the agent proposes:

  • Input/output entity structure based on your dataset.

image (10).png

 

What you check: ✅

  • Check if all input and output entities exist.

    image (15).png

  • Check if you need an explanation output.

    image (16).png

  • Check input and output entities types, e.g., “image link” is configured as IMAGE entity type or you can ask why some entities have a certain type.

    image (12).png

    • You may ask about entity types for your outputs and adjust them.

image (13).png

  • Set word limits for outputs.

    image (14).png

  • Check extracted data for input entities, e.g. check if the “image link” is extracted correctly.

    image (11).png

What you do next: Accept or ask the agent to fix.

Step 5: Defining quality criteria

Step 5: Defining quality criteria

Actions:

  1. Review agent's proposed quality criteria.

  2. Verify criteria cover:

    • Question type classification accuracy.

    • Question and answer relevance.

    • Question and answer clarity.

 

What the agent proposes:

  • Quality Requirements with pass/fail conditions.

  • Criteria aligned with your generation rules.

 

What you check: ✅

  • Check if proposed quality criteria are not redundant.

  • Request the agent to edit criteria if necessary.

  • Quality Requirements address classification accuracy.

Generated quality element clearly show pass/fail conditions.

image (17).png

 

What You Do Next: Accept or ask the agent to fix.

 

What the agent confirms:

✅ Quality Requirements configured.

These requirements will be used by LLM QA to automatically check expert work.
Experts must meet these standards for their annotations to pass.

Ready to proceed to Expert UI?
Step 6: Building annotation interface

Step 6: Building annotation interface

Actions:

  • Create user friendly UI for experts.

What the agent proposes

  • Project UI for experts.

image (19).png

 

What you check: ✅

  • The workflow is logical: the image appears next to the question and answer fields.

  • The UI is user-friendly and easy to navigate.

What you do next: Accept or ask the agent to fix the UI.
 

Step 7: Writing annotator guidelines

Step 7: Writing annotator guidelines

Actions:

  1. Write clear instructions for experts:

    • Workflow description.

    • Evaluation metrics and scoring rubrics.

    • Reasoning expectations.

  2. Add examples:

    • Pass example.

    • Fail example.

 

What the agent generates:

image (20).png

 

What you check: ✅

Guidelines clearly explain the selection task workflow and match the UI flow.

image (21).png

Question types are listed.

image (22).png

Examples show the evaluation and selection process.

image (23).png

Quality rubrics are provided.

image (24).png

 

What you do next: Accept or ask the agent to fix the guideline. Click “Proceed”.

Step 8: Self-check (test your configuration)

Step 8: Self-check (test your configuration)

Actions:

  1. Complete at least 1-2 tasks tasks yourself.

  2. Test the interface, guidelines, and quality criteria.

  3. Review LLM QA evaluations.

 

What you check:

  • Interface usability.

  • Task completeness.

  • LLM QA responses.

  • Guideline clarity.

 

What you do:

  1. Click "Proceed" to navigate to the Self-check step.

  2. The system shows you 10 sample tasks from your dataset.

    image (25).png

  3. Open the task to review the UI.

  4. Check that all inputs are extracted correctly in the provided samples: image displays properly, fields for question and answer are shown, etc.

  5. Complete the first task yourself based on the UI:

    • View the image.

    • Write question and answer.

image (26).png

  • Submit your answer by clicking on the ”Submit to LLM QA”.

  • LLM QA evaluation will take 30-60 sec.

  1. Verify that the LLM evaluation and comments align with the pass or fail criteria. If not, go to Project Configuration step and improve project settings.

    image (27).png

  2. You can download completed test samples with status “Ready for Review” by selecting them.

 

What you do next: Complete 2-3 more test tasks with different image types to ensure everything works correctly.

  • If not, go to the Project Configuration step and improve the settings.

  • If yes, go to the Contract section by clicking “Proceed”.

Step 9: Set pricing and launch

Step 9: Set pricing and launch

Actions:

Agent recommends you several options with suitable experts. Choose your audience.

image (31).png

Once the audience chosen, proceed to confirm cost estimates by clicking “Save”.

image (32).png

Select number of tasks, estimate task completion time, set the price for completing one task.

image (33).png

  • Click "Launch" to start labeling.

Step 10: Monitor results

Step 10: Monitor results

What you see:

  • Tasks sent for annotation.

  • Status for each task:

    • Finding Expert - the task is available on the tasks marketplace.

    • In Expert Work - an expert is working on the task.

    • Quality Check - LLM QA is checking the task.

    • Ready for Review - the task is waiting for your review.

Actions:

  1. Wait (experts are working on your tasks - evaluation and selection takes time). It can takes 30-60 min.

  2. Track item statuses in dashboard:

    • Finding Expert → In Expert Work → Quality Check → Ready for Review → Completed

  3. Review results with status "Ready for Review".

  4. Check LLM QA evaluations.

  5. Monitor progress metrics.

image (30).png

 

What you check: ✅

  • Results are coming in.

  • Each result has a selected query and reason.

  • Quality checks are passing.

Step 11: Do user quality calibration

Step 11: Do user quality calibration

Actions:

  1. Review results in "Ready for Review" status.

    • Click on a completed item to review it.

  2. Evaluate each Quality Criterion (Pass/Fail).

    • LLM QA evaluation on quality elements (the naming may vary depending on how you configured).

  3. Write reasoning for each evaluation.

  4. Compare your results with LLM QA assessments.

Check & provide quality feedback: ✅

Review LLM comments on each quality elements (Selection Accuracy on Selected Query and Reasoning Quality on Selection Reasoning).

  • In early stages, review as many tasks as you want.

  • The review results will be considered in the next project iteration.

  • You'll be asked to leave comments if you disagree with the LLM QA verdict.

image (28).png

image (29).png

  1. If satisfied Download results as JSON.

  2. If you want to improve quality, complete review and create new project version, Agent will help you to change the settings based on your quality feedback.

  3. The agent will guide you through the next iteration of your project configuration.

 

Best practices

  1. Start small: Test with 10-20 images before scaling to thousands.

  2. Do self check.

  3. Be specific about selection criteria: Clearly explain to the agent what makes your data "best" (relevant context, natural language, appropriate specificity).

  4. Provide example selections: Share examples of good/bad results with reasons to help the agent understand your expectations.

  5. Review early results: Pay special attention in early results to ensure selections follow your rules (not too vague, not too specific).

  6. Iterate on quality requirements: Use Quality Feedback to refine project settings.

Was this article helpful?

23 out of 28 found this helpful