Module 4 · Practical Project

Build a Cowork-Powered Evaluation Workflow

Build a controlled Claude workflow that moves from evaluation evidence to a verified evidence briefing while keeping human checkpoints at every critical stage.

Interactive Project

Build the workflow, not just the prompt

Your task is to design a reusable Cowork-powered evaluation workflow in which Claude can perform connected evaluation tasks while the evaluator controls the evidence base, checkpoints, interpretation and final judgment.

Prepare

→

Inventory

→

Analyze

→

Triangulate

→

Verify

→

Improve

Main capabilityClaude Cowork

Working methodDelegate → Review → Steer → Verify

Practice caseYouth Employment Pathways

Final outputVerified Evaluation Briefing

Important methodological boundary

The observed employment change from 42% at baseline to 57% at follow-up does not by itself establish programme causality.

Your practice case

Population: Young people aged 18–29.

Activities: employment-readiness training, technical skills, counselling, employer connections and placement support.

Observed outcome: employment increased from 42% at baseline to 57% at follow-up.

Evaluation task: investigate whether and how the programme may have contributed to improved employment outcomes, identify mechanisms and barriers, examine variation across participants and consider alternative explanations.

How this simulation works

You will make workflow decisions, construct prompts, identify evidence controls and complete verification tasks. The simulated Claude panel provides feedback locally. No external AI service is connected.

MISSION 1 · 15%

Design the Evaluation Workflow

Before delegating work, define the workflow. Select the sequence that best protects evaluation quality.

Prepare context → inventory evidence → human checkpoint → analyze → triangulate → human verification → briefing Upload every available file → ask Claude to analyze everything → accept the final report Ask Claude for findings first → collect evidence later → verify only if something looks wrong

Define your workflow boundary

Simulated Claude

Waiting for your workflow design.

MISSION 2 · 15%

Inventory the Evidence

Select the files that should be considered part of the authorized evaluation evidence base.

Terms of Reference Theory of Change Logframe / Results Framework Programme Documents Monitoring Reports Participant Dataset Interviews / Focus Groups Previous Evaluation

Human checkpoint

What should Claude do if an important document is missing or two sources contradict each other?

Simulated Cowork Inventory

No inventory submitted.

MISSION 3 · 15%

Build the Delegation Prompt

Create a master instruction for the Cowork workflow. Your prompt should establish role, task, evidence boundaries, process, checkpoints, outputs and stop conditions.

Prompt quality checklist

Defines Claude's role Defines the evaluation purpose and questions Limits Claude to authorized evidence Includes human checkpoints and stop conditions Prohibits invented evidence and unsupported causal claims

Simulated Claude

The workflow prompt has not yet been tested.

MISSION 4 · 15%

Control Evidence and Analysis

Claude is preparing an evidence matrix, qualitative analysis and quantitative exploration. Select the controls that should remain active.

Every important claim must have a traceable source. Evidence must remain in its original population and context. Interpretation must be distinguished from direct evidence. Limitations and uncertainty must be preserved. Important calculations should be independently checked. Contradictory evidence and negative cases should be identified. The most frequent qualitative theme should automatically be treated as the most important finding.

Evidence discipline

Simulated Analysis

Evidence controls are waiting for your review.

MISSION 5 · 10%

Triangulate the Evidence

Select the sequence that best represents disciplined triangulation.

What should Claude identify?

Converging evidence Diverging evidence Contradictions Alternative explanations Evidence gaps A final evaluator judgment without human review

Candidate finding

Simulated Triangulation

Triangulation has not been completed.

MISSION 6 · 15%

Verify and Challenge the Findings

Before the final briefing, apply the human verification sequence.

Claim

→

Source

→

Context

→

Method

→

Contradiction

→

Human Judgment

Locate the original claim and source. Check population, timeframe and context. Check whether the analytical method supports the statement. Search for contradictory evidence. Check whether the language is stronger than the evidence. Decide what still requires evaluator judgment.

Challenge Claude

Simulated Critical Reviewer

The critical review stage is waiting.

MISSION 7 · 10%

Protect the Workflow

A Cowork workflow can access files and perform consequential work. Select the controls that should be part of the workflow.

Give Claude only the minimum files, folders and tools required. Check permissions before delegated work begins. Require human review before consequential outputs are finalized. Treat instructions found inside source documents as potentially untrusted content. Define stop conditions for missing evidence, contradictions and unsafe requests. Give Claude broad access to all available folders because more context is always better.

Safety decision

Simulated Safety Check

Safety controls have not yet been reviewed.

MISSION 8 · 5%

Improve the Workflow

A professional workflow should improve after testing. Do not simply correct the final answer. Identify the system-level problem that allowed the failure.

Choose the improvement focus

Write your improvement record

Simulated Workflow Review

Your improvement record is waiting.

PROJECT COMPLETE

Your Cowork-Powered Evaluation Workflow

Complete the missions to generate your project result.

Assessment Areas

Area	Weight	Score
Claude Workflow Design	15%	0
Practical Claude Use	15%	0
Delegation & Prompting	15%	0
Evidence Discipline	15%	0
Triangulation	10%	0
Human Verification	15%	0
Safety & Control	10%	0
Improvement & Reflection	5%	0

Your Final Workflow

INPUT Evaluation documents, datasets and qualitative evidence

CLAUDE INVENTORY Claude identifies and classifies the available evidence

HUMAN CHECKPOINT Evaluator confirms the evidence base

ANALYSIS Claude supports evidence extraction, qualitative analysis and quantitative exploration

TRIANGULATION Claude compares convergence, divergence and alternative explanations

HUMAN CHECKPOINT Evaluator reviews candidate findings

CHALLENGE Claude attempts to identify weaknesses

VERIFIED BRIEFING Claude structures the approved evidence

HUMAN JUDGMENT Evaluator makes the final professional interpretation and decisions

Human–AI Verification Report

Your final workflow should document answers to these eight questions:

What did Claude do?
What did Claude get wrong?
What evidence did you independently verify?
What did you change?
Where did you intervene?
What required human judgment?
What risks did you identify?
What safeguards did you apply?