Build Your AI-Powered Evaluation System
Design, build, test and verify a reusable Claude-powered evaluation system that keeps evidence integrity, safety and human judgment at the center.
Build the system around the evaluator
Your challenge is to move from using Claude for individual tasks to designing a controlled professional system that integrates Claude across an evaluation workflow.
This project is not about automating an entire evaluation. It is about demonstrating where AI assistance is appropriate while preserving evidence integrity, safety, transparency, verification and human judgment.
Practice scenario
Youth Employment Pathways Programme
Target group: Young people aged 18–29.
Programme: Employment-readiness training, technical skills development, counselling and employer connections.
Observed outcome: Reported employment increased from 42% at baseline to 57% six months after completion.
Evaluation purpose: Investigate whether and how the programme may have contributed to employment outcomes, including mechanisms, barriers, variation and alternative explanations.
The increase from 42% to 57% is observed change. It does not, by itself, establish that the programme caused the change.
This tool does not connect to Claude or any external AI service. It simulates project decisions, checks your responses locally and provides feedback as you progress.
Define the Evaluation
Before introducing AI, define the evaluation itself. Select the elements that belong in your evaluation definition.
Write your evaluation definition
Build the AI Task Map
For each type of evaluation work, decide what role Claude should play. Select the tasks that are appropriate for AI assistance.
Delegation decision
Build the Claude Project
Your Claude Project should contain organized evaluation context, knowledge and instructions. Select the elements that belong in a professional Project.
Project instructions
Build the Prompt System
Create a reusable prompt system rather than a collection of unrelated prompts.
Professional prompt structure
Complete the structure:
Add Research
Identify one research question for which external evidence could strengthen the evaluation.
Research record
Claude's research summary is not itself the evidence. The underlying sources must be checked.
Create Evaluation Artifacts
Select at least three reusable evaluation artifacts for your system.
Artifact quality
Build the Repeatable Workflow
Connect the individual capabilities into a repeatable evaluation process.
Workflow Contract
Delegate One Bounded Task
Select the most appropriate bounded task for Cowork or an agentic workflow.
Delegation boundaries
Design the Safety Architecture
Select the safety controls that should be part of the system.
Safety principle
Test the System
A workflow is not ready simply because it works once. Select the tests that should be performed.
Choose the expected behavior for missing evidence
Verify AI Outputs
Apply the verification sequence to important AI-generated claims.
Select the elements that should be checked for at least five important claims.
Map Human Judgment
Identify decisions that must remain under the authority of the human evaluator.
Conduct the Final System Audit
Complete the final professional audit. Your system should satisfy all of the following.
Final professional decision
Your AI-Powered Evaluation System
Complete the project missions to generate your final result.
Project Assessment
| Competency | Weight | Score |
|---|---|---|
| Evaluation Definition | 8% | 0 |
| AI Task Map | 8% | 0 |
| Claude Project | 8% | 0 |
| Prompt System | 10% | 0 |
| Research | 7% | 0 |
| Evaluation Artifacts | 8% | 0 |
| Workflow Design | 10% | 0 |
| Delegated Task | 8% | 0 |
| Safety Architecture | 10% | 0 |
| Testing | 8% | 0 |
| Verification | 10% | 0 |
| Human Judgment | 8% | 0 |
| Final System Audit | 7% | 0 |
Human–AI Verification Report
Your completed system should document:
- What did Claude do?
- What did Claude get wrong?
- What evidence did you verify?
- What did you change?
- What required human judgment?
- What risks did you identify?
- What safeguards did you apply?
Final System Architecture
Evaluation purpose, questions, criteria, methodology and evidence.
Determine what Claude should assist with, analyze, research, challenge, organize or delegate.
Create the Claude Project, context, knowledge and prompt system.
Use appropriate evidence analysis and reusable artifacts.
Connect tasks into a repeatable process.
Bound delegated work and protect information and permissions.
Trace, check, challenge and independently verify important outputs.
Retain professional authority over interpretation, conclusions and recommendations.