Learn to Use Vibe Coding for Monitoring & Evaluation

A practical step-by-step tutorial for M&E, MEL, MEAL, research and international development professionals.

Vibe coding is changing who can build software.

You no longer need to be an experienced programmer to prototype a calculator, data-quality checker, dashboard, reporting utility or small interactive M&E application.

But there is an important distinction:

The real opportunity for M&E professionals is to combine your methodological knowledge, professional judgement and understanding of the problem with AI’s ability to generate and modify software.

This tutorial takes you through that process by building a small M&E Indicator Quality Checker from scratch.

The central principle is simple:

What You Will Learn

• What vibe coding means.
• Which M&E tasks are suitable for vibe coding.
• How to turn an M&E problem into a software specification.
• How to write effective coding prompts.
• How to build a simple browser-based M&E tool.
• How to test AI-generated code.
• How to identify methodological and technical errors.
• How to challenge AI-generated applications.
• How to document assumptions and limitations.
• How to consider privacy and responsible use.
• How to distinguish a prototype from a production system.
• How to decide when an AI-built tool is ready for use.

1. What Is Vibe Coding?

Vibe coding is an approach to software development where you describe what you want an application to do in natural language and an AI coding system generates or modifies much of the underlying code.

Instead of manually writing every line of HTML, CSS, JavaScript, Python or another programming language, you communicate with the AI conversationally.

For example:

Build a mobile-friendly web application for M&E professionals.
The user should enter:
- indicator statement
- result level
- unit of measurement
- baseline
- target
- data source
- reporting frequency
- disaggregation
The application should identify missing information and potential issues.
Do not make definitive methodological judgements.

The AI can create a first version. You can then ask it to add validation, correct calculations, improve the interface or change specific rules.

That iterative process is the useful part of vibe coding.

2. Why Vibe Coding Matters for M&E

M&E professionals regularly perform structured and repetitive tasks that can sometimes be turned into small digital tools.

• Indicator achievement calculations.
• Data-quality checks.
• Results-framework reviews.
• Survey validation.
• Sample-size calculations.
• Evidence matrices.
• Recommendation tracking.
• Reporting completeness checks.
• Simple monitoring dashboards.
• Coverage calculations.
• Budget and cost-efficiency calculations.
• Document and evidence organization.

The objective is not to turn every M&E process into software.

The objective is to identify small, repeatable processes where a simple application could reduce manual work without replacing professional judgement.

3. Which M&E Tasks Are Good Candidates?

4. Choose a Small First Project

Your first project should be:

• small;
• clearly defined;
• based on rules you understand;
• easy to test;
• low-risk;
• useful to a real user;
• preferably capable of running without a backend.

Do not start with:

Start with:

5. The Practical Project: M&E Indicator Quality Checker

The user will enter:

1. Indicator statement
2. Result level
3. Unit of measurement
4. Baseline
5. Target
6. Data source
7. Reporting frequency
8. Disaggregation

The application will identify missing information and potential issues.

A useful output might say:

• Clarity: Review recommended.
• Measurement: Review recommended.
• Baseline: Information provided.
• Target: Information provided.
• Data source: Information provided.
• Frequency: Information provided.
• Disaggregation: Information provided.

The application should support professional review, not pretend to replace it.

6. Step 1 — Define the M&E Problem Before Writing Code

Do not begin with “Build me an application.”

Begin with the problem.

I want to build a small web tool for M&E professionals that helps them review the completeness and potential quality issues of an indicator.
Before writing any code, help me define:
1. The problem
2. Intended users
3. Workflow
4. Inputs
5. Outputs
6. Risks
Do not generate code yet.

Read the proposed design before moving forward.

7. Step 2 — Define the Users

The primary users are M&E officers, MEL officers, programme managers and evaluators.
Assume that users understand M&E concepts but may have little or no programming experience.
Use clear professional language and do not assume technical knowledge.

This matters because a tool designed for software developers may use workflows and terminology that are inappropriate for M&E professionals.

8. Step 3 — Define the Inputs

The user should enter:
1. Indicator statement
2. Result level
3. Unit of measurement
4. Baseline
5. Target
6. Data source
7. Reporting frequency
8. Disaggregation
Identify the appropriate field type for each input and explain why.

Review the AI’s choices. Do not accept them automatically.

9. Step 4 — Define the Methodological Rules

This is where your M&E expertise becomes essential.

The tool should check for:
- missing indicator statement
- potentially vague wording
- missing unit of measurement
- missing baseline
- missing target
- missing data source
- missing reporting frequency
- missing consideration of disaggregation
It should identify potential issues rather than make definitive professional judgements.
For each check:
1. Define the rule.
2. Explain the rationale.
3. Identify whether it is objectively testable.
4. Identify where professional judgement is required.

This prevents the AI from silently turning its assumptions into methodology.

10. Step 5 — Do Not Let AI Invent Methodology

Suppose the AI says:

You may disagree.

Respond with a controlled correction:

Do not assume that every indicator must have a baseline.
Flag a missing baseline as something to review.
Explain that the appropriateness of a baseline depends on the indicator and measurement design.
Do not change any other assessment logic.

11. Step 6 — Give AI Good and Weak Examples

Concrete examples help AI understand the behaviour you expect.

GOOD EXAMPLE
Indicator:
Percentage of programme participants who report increased household income six months after programme participation.
Result level:
Outcome
Unit:
Percentage of participants
Baseline:
42%
Target:
65%
Data source:
Participant survey
Frequency:
Annual
Disaggregation:
Sex, age and location
Explain how the tool should assess this example.

WEAK EXAMPLE
Indicator:
Improved capacity.
Result level:
Outcome
Unit:
Not specified
Baseline:
Not specified
Target:
Not specified
Data source:
Not specified
Frequency:
Not specified
Disaggregation:
Not specified
Explain what the tool should flag.

12. Step 7 — Define the Output

After clicking "Review Indicator", display:
1. Overall review summary
2. Individual criteria
3. Status
4. Explanation
5. Suggested action
Use three statuses:
- Complete
- Review
- Missing
Do not use a numerical "quality score".
Do not create false precision.

A score such as 82/100 may look scientific without a validated scoring methodology.

13. Step 8 — Ask AI to Design the Application

Based on the approved requirements, design the first version.
The application should:
- run in a web browser
- use HTML, CSS and JavaScript
- require no backend
- require no external API
- process entered information locally
- work on desktop and mobile
- use accessible form controls
- use clear professional M&E language
Before writing code, provide a short implementation plan.

If the architecture becomes unnecessarily complicated, ask:

Simplify the architecture.
This is a prototype.
Avoid unnecessary frameworks, databases, authentication and dependencies.

14. Step 9 — Generate the First Code

Implement the first working version based on the approved specification.
Keep the code simple and well structured.
Include comments explaining the main assessment rules.
Do not add features that were not specified.

15. Step 10 — Test the Application

Start with deliberately simple cases.

Test 1:
Improved capacity
Everything else blank.
Expected:
Multiple potential issues are identified.

Test 2:
Percentage of women participants who report increased use of climate-resilient agricultural practices six months after programme support
Complete supporting information.
Expected:
The application produces a reasonable review without making unsupported judgements.

16. Step 11 — Build a Test Matrix

Create your own tests. Do not rely only on AI-generated tests.

17. Step 12 — Ask AI to Attack Its Own Application

Now use AI as a critical reviewer rather than only as a builder.

Act as a hostile software tester and an experienced M&E reviewer.
Try to find ways this application could produce incorrect, misleading or confusing results.
Review:
- methodological assumptions
- validation logic
- calculations
- user interface
- accessibility
- edge cases
- privacy
- misleading language
- false precision
Do not modify the application yet.
First provide a structured risk report.

You are now asking AI to challenge what it created.

18. Step 13 — Fix One Problem at a Time

Suppose the AI identifies this issue:

Make a controlled change:

Correct this behaviour.
A missing disaggregation should generate a "Review" message rather than a "Missing" error.
Explain that the user should consider whether relevant disaggregation is appropriate for the indicator and evaluation purpose.
Do not change any other assessment logic.

This is much safer than asking the AI to “fix the app” without specifying the intended change.

19. Step 14 — Document Every Important Rule

List every methodological rule currently implemented.
For each rule provide:
- rule description
- why it exists
- what triggers it
- message shown to the user
- limitations
- whether human judgement is required

This becomes part of your methodological documentation and helps expose assumptions introduced by AI.

20. Step 15 — Test With Realistic M&E Scenarios

Take three to five indicators from previous projects, after removing confidential and personally identifiable information, and test the application.

• Did the tool identify useful issues?
• Did it flag anything incorrectly?
• Did it miss something important?
• Was the language understandable?
• Would an M&E officer find the output useful?
• Did it imply more certainty than the evidence supports?

This tests the professional usefulness of the tool, not just whether the buttons work.

21. Step 16 — Improve the Interface After the Logic Works

Improve the user interface while keeping all existing assessment logic unchanged.
Make the workflow clearer for an M&E professional.
Make results easy to scan.
Clearly distinguish between:
- complete information
- missing information
- areas requiring professional review
Optimise for mobile devices.
Do not add unnecessary animations, pop-ups or decorative elements.

Methodology first. Interface second.

22. Step 17 — Add a “How This Tool Works” Section

Every professional tool should explain its boundaries.

Ask AI to document:

• What the tool checks.
• What it does not check.
• How users should interpret the results.
• Which rules are objective.
• Which issues require professional judgement.
• Why the tool does not provide a numerical quality score.
• What limitations remain.

23. Step 18 — Review Privacy and Data Protection

Do not assume that a browser-based tool is automatically private.

Ask the AI to inspect the actual implementation:

Review this application from a data-protection perspective.
Identify whether user-entered information is transmitted outside the browser.
Identify external dependencies, APIs, analytics, storage mechanisms and third-party services.
Explain the risks of entering confidential monitoring or evaluation information.
Do not assume that local processing is guaranteed. Verify the implementation.

Pay particular attention to:

• Beneficiary information.
• Survey responses.
• Interview data.
• Health information.
• Personal identifiers.
• Geographic information.
• Protection or case-management information.
• Confidential programme data.

This approach is consistent with broader responsible-AI frameworks that emphasise privacy, transparency, security, accountability and human oversight. OECD AI Principles, UNESCO Recommendation on the Ethics of AI and the NIST AI Risk Management Framework are useful references.

24. Step 19 — Perform a Final Acceptance Test

Perform a final review of this application.
Check whether it meets the original specification.
Create a table with:
- Requirement
- Implemented?
- Evidence in the application
- Potential issue
- Recommended action
Do not add new features.

This gives you a simple acceptance-test process before the tool is used more widely.

25. Step 20 — Decide Whether It Is Ready

26. Practical Daily Workflow for M&E Professionals

Vibe coding becomes more useful when you connect it to work you already do.

A useful rule is:

27. From Prototype to Real M&E Tool

Not every AI-built application has the same risk.

A prototype working on your laptop is not automatically a production-ready application.

28. What Vibe Coding Should Not Replace

• Methodological expertise.
• Statistical knowledge.
• Professional judgement.
• Data protection.
• Ethical review.
• Software testing.
• Security review.
• Accessibility testing.
• User testing.
• Formal software engineering for complex systems.

The more consequential the application, the stronger these controls need to become.

29. A Reusable Vibe-Coding Prompt for M&E

PROJECT
I want to build a small digital tool for Monitoring & Evaluation professionals.
PROBLEM
[Describe the M&E problem.]
USERS
[Describe the intended users.]
INPUTS
[List what users enter or upload.]
OUTPUTS
[Describe what the application should produce.]
METHODOLOGICAL RULES
[Define the rules the application must follow.]
IMPORTANT ASSUMPTIONS
Do not invent methodological rules that I have not specified.
Identify ambiguities and ask for clarification where necessary.
USER EXPERIENCE
Make the application simple, accessible, mobile-friendly and appropriate for professional M&E users.
PRIVACY
Do not send user data to external services unless explicitly required and approved.
TESTING
Create test cases for normal, missing, invalid, extreme and ambiguous inputs.
FIRST STEP
Do not write code yet.
First analyse:
1. User workflow
2. Inputs
3. Outputs
4. Methodological rules
5. Validation rules
6. Edge cases
7. Privacy considerations
8. Testing strategy
9. Acceptance criteria
Wait for my approval before generating code.

30. What Else Can You Build?

• M&E Indicator Calculator: achievement, variance, percentage change and target progress.
• Indicator Quality Checker: clarity, measurability, baseline, target, data source and frequency.
• M&E Data Quality Checker: missing values, duplicates, inconsistent categories, invalid dates and potential outliers.
• Results Framework Checker: potential gaps in results, indicators, assumptions, targets and responsibilities.
• Evaluation Evidence Matrix: evaluation questions, evidence sources, findings, confidence and limitations.
• Recommendation Tracker: recommendation owner, status, deadline and evidence of follow-up.
• Reporting Completeness Checker: identify missing sections, indicators, annexes and supporting evidence.
• M&E Dashboard Prototype: targets, actuals, trends, geography, disaggregation, risks and data-quality flags.

31. The Most Important Question

That question is more important than asking whether the AI produced impressive code.

Your professional value is knowing:

• what the tool should do;
• what it should not do;
• which assumptions are acceptable;
• what constitutes an error;
• how the output should be tested;
• when human judgement is required;
• when the application is safe enough to use.

32. Vibe Coding Changes the Role of the M&E Professional

This shifts the M&E professional closer to the design of digital tools without eliminating the need for software engineering where systems become complex or high-risk.

33. Your First Challenge

Choose one repetitive task that currently requires:

• Excel;
• manual calculations;
• copying and pasting;
• repeated checking;
• manual formatting;
• repetitive reporting.

Then ask:

If yes, you may already have the beginning of your first vibe-coded M&E tool.

Further EvalCommunity Resources

Continue building your AI capability with these related EvalCommunity resources:

• Claude Cowork for Monitoring & Evaluation — practical task automation for recurring M&E workflows.
• Five Practical Lessons for Using AI Agents — decision boundaries, verification, tripwires and human oversight.
• Start from Zero with Agentic AI — a practical roadmap for M&E and MEAL professionals.
• Data Quality Validation for Your AI Agent — building validation into AI-enabled M&E workflows.
• Human-First AI Manifesto for M&E — principles for keeping human judgement, ethics and accountability at the centre.

Authoritative AI Governance Resources

• OECD AI Principles
• UNESCO Recommendation on the Ethics of Artificial Intelligence
• NIST AI Risk Management Framework
• NIST AI RMF Playbook
• Anthropic: Use Claude Cowork Safely

Continue Learning With EvalCommunity Academy

Vibe coding is one part of a broader shift toward AI-assisted M&E. The deeper skill is learning how to combine AI capability with methodological reasoning, evidence quality, testing, ethics and human oversight.

The AI in Monitoring & Evaluation Certificate provides practical training across AI foundations, evaluation design, data collection, qualitative and quantitative analysis, reporting, governance and responsible AI use.

For professionals who want to move further into automation and AI-enabled workflows, the AI for M&E Professional Bundle combines the AI in M&E Certificate with the AI Agents for Evaluators Certificate.

Start small. Define the methodology. Build quickly. Test aggressively. And never outsource professional judgement to the code.