Task Actions

The Actions button on each task is your gateway to Empromptu's optimization capabilities. This is where you access all the tools needed to improve your AI application's performance.

What you'll learn ⏱️ 5 minutes

  • What happens when you click the Actions button

  • Overview of all available optimization tools

  • When to use each action type

  • How actions work together to improve performance

  • Best practices for task optimization

Accessing Task Actions

From your project dashboard, each task has an Actions button in the rightmost column. Click this button to access optimization tools for that specific task.

Example Task Row:

Task Name
Description
Initial Accuracy
Current Accuracy
Status
Actions

Review Summarizer

This task summarizes reviews we scraped off the web

50%

65%

Active

Actions ← Click here

When you click Actions, you'll see five optimization options available for that task.

Available Actions

1. Prompt Optimization

What it does: Creates and manages Prompt Families to improve response quality and accuracy.

Key capabilities:

  • Build Prompt Families (collections of specialized prompts)

  • Run automatic optimization to improve performance

  • Manually refine prompts for specific scenarios

  • View optimization history and results

When to use:

  • Your task has low accuracy scores

  • Responses are inconsistent across different inputs

  • You want to achieve 90%+ accuracy

  • Initial setup after creating a new task

What you'll find: Event Log, Prompt Family management, Manual Optimization wizard, Automatic Optimization

2. Input Optimization

What it does: Manages test data and analyzes real user inputs to improve task performance.

Key capabilities:

  • Create manual test inputs for optimization

  • Monitor real end-user inputs and performance

  • Analyze input patterns and edge cases

  • Use input data to guide optimization

When to use:

  • Before running optimization (need test data)

  • After deployment (analyze real user behavior)

  • When discovering new edge cases

  • To understand what inputs cause problems

What you'll find: Manual Inputs creation, End User Inputs analytics, input performance tracking

3. Model Optimization

What it does: Tests different AI models to find the best fit for your specific use case.

Key capabilities:

  • Compare performance across different models

  • Test models like GPT-4o, Claude 3 Opus, Claude 3 Sonnet

  • Adjust temperature and other parameters

  • Analyze cost vs performance trade-offs

When to use:

  • Task performance plateaus with current model

  • Cost optimization is needed

  • Specific model features are required

  • Initial setup to find optimal model

What you'll find: Model selection interface, side-by-side comparisons, parameter tuning options

4. Edge Case Detection

What it does: Identifies problematic inputs using visual analysis and helps resolve them.

Key capabilities:

  • Visual scatter plot of task performance

  • Identify clusters of low-performing inputs

  • Select problem areas for targeted optimization

  • Understand performance patterns

When to use:

  • After running initial optimization (need data points)

  • When overall accuracy is good but some inputs fail

  • To find patterns in problematic scenarios

  • For systematic problem identification

What you'll find: Performance scatter plot, score clustering visualization, optimization targeting tools

5. Evaluations

What it does: Defines success criteria that guide optimization and measure performance.

Key capabilities:

  • Create custom evaluation criteria

  • Use automatic evaluation generation

  • Manage active and inactive evaluations

  • Set specific quality standards

When to use:

  • Before any optimization (define success first)

  • When optimization isn't targeting the right goals

  • To add new quality requirements

  • When business requirements change

What you'll find: Evaluation creation tools, criteria management, success metrics definition

How Actions Work Together

1. Start with Evaluations Define what success looks like before optimizing:

Actions → Evaluations → Create success criteria

2. Add Input Data Provide test data for optimization:

Actions → Input Optimization → Add manual inputs

3. Run Initial Optimization Use automatic optimization to establish baseline:

Actions → Prompt Optimization → Automatic Optimization

4. Find and Fix Problems Use visual tools to identify issues:

Actions → Edge Case Detection → Target problem areas

5. Fine-tune Performance Test different models and manual optimization:

Actions → Model Optimization → Compare modelsActions → Prompt Optimization → Manual refinement

Action Interdependencies

Actions That Build on Each Other:

Evaluations → All Other Actions

  • Evaluations define success criteria for all optimization

  • Must be set up before meaningful optimization can occur

Input Optimization → Edge Case Detection

  • Need input data to generate the scatter plot visualization

  • More inputs create better problem identification

Prompt Optimization → Model Optimization

  • Different models may work better with different prompt styles

  • Optimize prompts first, then test models

Edge Case Detection → Prompt Optimization

  • Identifies specific problems for targeted prompt improvement

  • Creates focused optimization goals

Best Practices by Action Type

Prompt Optimization Best Practices:

  • Start with automatic optimization for baseline performance

  • Use manual optimization for specific problem areas

  • Build diverse Prompt Families for different input types

  • Monitor Event Log to understand what works

Input Optimization Best Practices:

  • Create representative manual inputs that match real use cases

  • Include edge cases and difficult scenarios

  • Monitor end-user inputs after deployment

  • Add new inputs when discovering problems

Model Optimization Best Practices:

  • Test models after prompt optimization is complete

  • Consider cost vs performance trade-offs

  • Adjust temperature based on use case (creative vs precise)

  • Document which models work best for which scenarios

Edge Case Detection Best Practices:

  • Wait until you have sufficient data points (15-20+ optimization runs)

  • Focus on clusters of low-performing inputs

  • Use targeted optimization for identified problem areas

  • Regular review as new data comes in

Evaluations Best Practices:

  • Start with 3-5 core evaluations covering key requirements

  • Use specific, measurable criteria

  • Balance automatic and manual evaluation creation

  • Update evaluations as requirements evolve

Common Action Workflows

New Task Setup:

  1. Evaluations → Define success criteria

  2. Input Optimization → Add test inputs

  3. Prompt Optimization → Run automatic optimization

  4. Review results → Check if performance meets goals

Performance Improvement:

  1. Edge Case Detection → Identify problem areas

  2. Input Optimization → Add problematic inputs as test cases

  3. Prompt Optimization → Manual optimization for specific issues

  4. Model Optimization → Test if different model performs better

Production Monitoring:

  1. Input Optimization → Monitor end-user inputs

  2. Edge Case Detection → Find new problem patterns

  3. Evaluations → Add criteria for new requirements

  4. Prompt Optimization → Continuous improvement based on real usage

Troubleshooting Actions

Actions Button Not Working:

Check: Task status is "Active" Solution: Inactive tasks can't be optimized

Limited Action Options:

Check: Task has been used/optimized before Solution: Some actions require initial data or optimization

Poor Optimization Results:

Check: Evaluations are well-defined and inputs are representative Solution: Improve evaluation criteria and add better test inputs

Can't Access Certain Features:

Check: Sufficient optimization runs have occurred Solution: Edge Case Detection requires 15-20+ runs to unlock

Understanding Action Results

Success Indicators:

  • Accuracy scores improving (5.0 → 7.5 → 8.9)

  • More consistent performance across different inputs

  • Better real-world user feedback after deployment

  • Fewer edge cases identified in detection analysis

Warning Signs:

  • Scores plateauing despite optimization attempts

  • High variance in performance across similar inputs

  • New problems appearing with different input types

  • End-user performance declining after changes

Next Steps

Now that you understand Task Actions:

  • Learn Prompt Optimization: Master the core optimization technology

  • Understand Evaluations: Set up effective success criteria

  • Explore Edge Case Detection: Use visual tools to find problems

  • Check Input Optimization: Manage test data and real user analytics

Last updated