Before Using Claude Cowork for M&E, Build This Workspace
Practical AI Tutorial for M&E and International Development
Before You Use Claude Cowork for M&E, Build This Workspace First
Create a controlled workspace that helps Claude organise evidence, follow programme rules, protect sensitive information and produce stronger drafts for evaluation, reporting, learning and development decision-making.
Quick answer
Claude Cowork becomes more reliable when it does not start from an unstructured folder. Give it separate areas for professional context, operating rules, active projects and generated outputs. Add a project brief, source register, risk check, evidence ledger and human-review step before using any result.
What you will build
Separate rules, context, projects and outputs.
Define your sector context and Claude’s boundaries.
Track source status, uncertainty and contradictions.
Verify data, ethics, interpretation and approval.
What Claude Cowork changes
In a normal chat, the prompt carries most of the task context. In Cowork, Claude can work across the folders, files, project instructions and tools that you make available. On desktop, local file access depends on the folders you connect and the permissions you grant. Cowork can also organise related work into projects with their own files, context, instructions and memory.
This makes multi-file work easier, but it also increases the importance of access control. Claude may be able to read, create, edit or delete files inside the permitted workspace. Use a dedicated working folder, keep backups and avoid granting broad access to unrelated or sensitive records.
Why this matters in M&E and development work
A weak assumption in an ordinary office task may create an inconvenient edit. In evaluation, humanitarian or development work, it can misrepresent results, expose participants, distort an indicator, erase country differences, overstate contribution or create an unsupported recommendation that influences resources and decisions.
Why one long prompt is not enough
A prompt such as the following describes the task, but not the rules governing the task:
Claude still does not know which results framework is current, whether a partner narrative is approved, how indicator periods should be compared, which quotation may be published, how conflicting evidence should be handled or whether the output is intended for a donor, programme team or community audience.
A workspace moves stable expectations into reusable files. The task prompt then activates the system instead of repeating every rule.
The four-folder workspace
Your role, programmes, audiences and terminology
Evidence, quality, ethics and file-handling rules
Brief, source register, data, evidence and templates
Drafts, logs, reviewer questions and approved versions
M&E-COWORK-WORKSPACE/ | |-- context/ |-- system/ |-- projects/ `-- outputs/
1. context/
Stores stable information about your professional role, sectors, programmes, audiences, preferred terminology, recurring frameworks and source priorities.
2. system/
Stores rules for evidence, data protection, inclusion, uncertainty, contradictions, calculations, approvals and prohibited assumptions.
3. projects/
Contains one controlled folder for every evaluation, reporting cycle, evidence review, programme design task, learning process or management decision.
4. outputs/
Keeps AI-produced drafts, evidence ledgers and review questions separate from original evidence. Use clear version labels and never overwrite sources.
Create a dedicated working copy
Do not connect Cowork directly to a master archive, case-management folder or unrestricted shared drive. Copy only the files required for the task into a dedicated workspace, remove unnecessary identifiers, keep backups and save generated material separately.
The core files that make the workspace useful
Start with two stable files and three project controls. Add more only when the work demonstrates a repeated need.
Who you are, what you produce and how you communicate.
How Claude must handle evidence, risk and uncertainty.
Purpose, audience, questions, boundaries and outputs.
File status, authority, period, owner and restrictions.
Claim-to-source traceability, limitations and confidence.
File 1: describe how you actually work
Your context file should reflect real responsibilities and recurring work, not an idealised profile. Include programme types, geographic scope, donor context, technical areas, languages, audiences and the distinctions you expect Claude to preserve.
# My M&E and International Development Work ## My role I work as a Monitoring, Evaluation and Learning adviser supporting international development and humanitarian programmes. My work includes: - results-framework and indicator design - evaluation planning and evidence synthesis - quantitative and qualitative analysis - data-quality review - donor and management reporting - outcome harvesting and learning - programme adaptation and follow-up ## Programme context I support multi-country programmes working through local partners. Country differences must remain visible. Do not combine findings across countries unless the data, definitions and periods are comparable. ## Typical deliverables - evaluation matrices and inception inputs - monitoring and donor reports - indicator performance reviews - evidence tables and learning briefs - qualitative coding frameworks - management decision briefs - dashboards and presentation outlines ## Main audiences Senior leaders need concise implications and decisions. Technical teams need methods, evidence and limitations. Donors need alignment with approved indicators and commitments. Partners need clear, respectful and actionable feedback. Community-facing materials need accessible language and privacy protection. ## Communication style Use clear professional language. Separate evidence, interpretation, judgement and recommendation. Avoid promotional language and exaggerated claims. Explain technical terms when the audience may not know M&E terminology. ## Terminology Use “programme participant” instead of “beneficiary” unless required. Do not use “impact” for activities, outputs or short-term outcomes. Distinguish reach, participation, satisfaction, output, outcome and impact. ## Source hierarchy Use the hierarchy specified in each project brief. Normally prioritise approved agreements, current results frameworks, validated datasets and approved guidance over unapproved drafts. Do not silently resolve conflicts. Record them for verification. ## Inclusion and context Do not infer gender, disability, ethnicity, vulnerability or social position. Use disaggregation only when it exists in the source data. Preserve the language and meaning of local stakeholders.
File 2: define how Claude must behave
The system file should control analysis, not only writing style. Include stop conditions that tell Claude when it must pause rather than improvise.
# M&E Cowork System Rules ## 1. Read before acting Read the context file, system rules, project brief, source register and required evidence before drafting. ## 2. Do not invent information Never invent findings, quotations, indicator values, baselines, targets, dates, budgets, sample sizes, locations, approvals or causal explanations. ## 3. Separate analytical layers Distinguish: - documented evidence - respondent or stakeholder perception - calculated result - analytical interpretation - evaluative judgement - recommendation - unresolved question ## 4. Preserve traceability Link every material claim to a source location where possible. Do not create citations to files or passages that are absent. ## 5. Report contradictions Identify both sources, describe the disagreement and state what must be verified. Do not silently choose the more convenient figure. ## 6. Handle uncertainty honestly State when evidence is incomplete, outdated, incomparable, translated, self-reported, unverified or based on a small or biased sample. ## 7. Protect people and sensitive information Do not reproduce personally identifiable or protection-sensitive details unless the brief explicitly authorises their use. ## 8. Preserve country and stakeholder differences Do not merge countries, institutions or participant groups when contexts, definitions, samples or reporting periods differ. ## 9. Validate quantitative claims Check numerator, denominator, unit, period, baseline, target, direction, disaggregation, cumulative status and calculation method. ## 10. Validate qualitative claims Preserve source codes, distinguish frequency from significance, identify contradictory cases and do not treat silence as evidence of absence. ## 11. Apply stop conditions Pause and ask before proceeding when: - the current results framework cannot be identified - reporting periods are incompatible - a required denominator is missing - sensitive data appears unauthorised - source status is unclear - the requested conclusion exceeds the available evidence ## 12. Preserve source files Do not overwrite original evidence. Save generated work in outputs/ and label it as a draft. ## 13. Require human review Do not finalise evaluation ratings, causal claims, safeguarding decisions, partner judgements, official statistics or donor submissions. ## 14. Complete a final check Review factual consistency, traceability, calculations, missing evidence, contradictions, confidentiality, inclusion, audience fit and unsupported claims.
Build a stronger project brief
A useful brief acts as an output contract. It explains not only what to produce, but what decision the output supports and what must remain outside the AI’s authority.
# Project Brief ## Purpose Prepare an indicator performance review for the quarterly programme meeting. ## Decision supported Help managers identify which results need verification, adaptation or partner follow-up. ## Audience Programme manager, technical leads and country coordinators. ## Reporting cut-off Use evidence available up to 30 June 2026. ## Required questions - Which indicators are on track, off track or not assessable? - Which data-quality issues could change the interpretation? - Which country or partner differences require attention? - What must be verified before management decides? ## Required sources Use the current approved results framework and validated reporting dataset. Treat partner narrative reports as explanatory evidence, not verified results. ## Boundaries - Do not infer reasons for underperformance. - Do not combine incompatible reporting periods. - Do not change approved indicator definitions. - Do not rank partners unless explicitly authorised. ## Outputs 1. Indicator status report. 2. Data-quality issues register. 3. Indicators requiring attention. 4. Partner follow-up questions. 5. Evidence and assumption log. ## Review owner MEL Manager. ## Output location outputs/indicator-review/v1-draft/
Add a source register without a wide mobile table
Record the status and restrictions of each important file. The card layout below remains readable on narrow screens and avoids forcing the whole page to scroll horizontally.
Status: Approved
Use: Indicator definitions and targets
Period: 2026
Restriction: Internal
Status: Validated
Use: Current indicator values
Period: April–June 2026
Restriction: Internal
Status: Draft
Use: Explanation of progress
Period: April–June 2026
Restriction: Not approved
Status: Raw qualitative evidence
Use: Stakeholder perspective
Period: June 2026
Restriction: Confidential
Add an evidence ledger and assumption log
The source register describes files. The evidence ledger describes the relationship between evidence and claims. This is especially useful for evaluation findings, donor narratives and management recommendations.
What statement is being tested?
Which file, sheet, row, section or interview code supports it?
What weakens, narrows or qualifies the evidence?
High, moderate, low or not assessable, with a reason.
What has not been verified?
Confirm, revise, reject or request more evidence.
Run a risk gate before giving Claude access
Not every task should use the same permissions or evidence. Classify the work before connecting a folder.
Lower risk
Formatting a public report, organising non-sensitive references, creating a blank template or summarising published guidance. Use ordinary review.
Moderate risk
Analysing de-identified monitoring data, drafting an internal donor report or coding qualitative evidence. Use minimum access, traceability and technical review.
High risk
Safeguarding cases, identifiable health or protection data, complaints, partner ratings, assistance decisions or security-sensitive records. Do not proceed without explicit organisational authorisation and appropriate controls.
Six questions before access
- Is every file necessary for this task?
- Can identifiers be removed or replaced with codes?
- Can an aggregate table replace raw records?
- Is the use permitted by organisational, donor and legal requirements?
- Who is responsible for reviewing the output?
- What files may Claude create, modify or delete?
The complete working method
Decide whether the data and task are suitable
Context, rules, brief, register and evidence
Resolve material gaps, conflicts and permissions
Create the matrix, report, brief or review
Check sources, calculations, gaps and contradictions
Revise, approve, reject or request more evidence
Eight practical M&E and development workflows
The workspace pattern remains stable. The project evidence, review rules and output contract change with the assignment.
1. Evaluation evidence matrix
Organise evidence by evaluation question before drafting findings or applying evaluation criteria.
2. Donor progress report
Draft against approved indicators and commitments without confusing activity completion with outcome achievement.
3. Outcome-harvesting analysis
Structure change narratives by actor, outcome level, country, institution, contribution and substantiation.
4. Indicator and early-warning review
Compare current values with targets while identifying missing, late, inconsistent or incomparable data.
5. Qualitative coding review
Support thematic analysis while protecting quotations, source codes and contradictory cases.
6. Data-quality assessment
Review completeness, validity, consistency, timeliness, uniqueness and integrity before analysis.
7. Theory of Change review
Test causal pathways, assumptions, external factors and evidence gaps without inventing programme logic.
8. Management decision brief
Turn verified evidence into options, implications and questions without allowing Claude to make the decision.
Protect country context and local meaning
International development work frequently combines evidence from different countries, institutions, languages and implementation models. Claude should not flatten these differences into one apparently consistent story.
Analyse countries separately before producing cross-country patterns.
Label translated quotations and record uncertainty or alternative meanings.
Distinguish government, partner, participant and programme-team perspectives.
Identify whose experience is absent rather than assuming silence means agreement.
Do not infer characteristics or create categories missing from the data.
Ask local reviewers to check cultural, political and institutional meaning.
Use a question-first workflow
Claude should ask questions only when the missing information could change the accuracy, safety or usefulness of the result. This is especially important when the source version, reporting period, intended audience, decision or permission is unclear.
Read the workspace context, system rules and relevant project folder first. Do not begin the deliverable immediately. Identify missing information, conflicting instructions, unclear permissions or uncertain source status that could materially affect the result. Ask only the questions needed to resolve those issues. After the questions are answered, create the requested draft, evidence ledger and reviewer questions in the specified outputs folder.
Use shorter prompts after the system is established
A short prompt becomes useful only after the workspace contains reliable context and rules.
Build the two reusable files through an interview
A guided interview can help you document your actual work patterns, terminology, evidence standards and review boundaries.
You are helping me create two reusable files for an AI-assisted M&E and international development workspace: 1. context/MY-M&E-WORK.md 2. system/M&E-SYSTEM-RULES.md Interview me one question at a time. Understand my role, programme context, recurring deliverables, audiences, terminology, source hierarchy, evidence standards, data-protection rules, country context, approval responsibilities and situations where AI must stop. Ask for concrete examples when my answer is vague. Distinguish real practice from aspiration. Identify contradictions between my answers. Do not use praise or conversational filler. Stop when you have enough information. Then generate both files and add a final section called: [Missing or weak information] Begin by asking about my current professional role and the programmes I support.
Apply output-specific quality checks
Quantitative evidence
- Check numerator and denominator.
- Confirm unit and reporting period.
- Separate cumulative and period values.
- Verify baseline, target and direction.
- Check missing, duplicate and outlier records.
- Confirm disaggregation categories and totals.
Qualitative evidence
- Preserve interview or document codes.
- Separate frequency from significance.
- Include contradictory and deviant cases.
- Record translation and interpretation limits.
- Check whether quotations could identify people.
- Identify absent stakeholder groups.
Evaluation judgement
- Use the criteria specified in the terms of reference.
- Show the evidence behind each judgement.
- Consider alternative explanations.
- State confidence and limitations.
- Do not invent ratings or scoring thresholds.
- Keep recommendations connected to findings.
Donor and management reporting
- Use the current approved template.
- Match the reporting cut-off.
- Separate verified results from narrative explanation.
- Flag missing evidence and approvals.
- Avoid unsupported attribution.
- Identify decisions requiring authority.
What Claude should not finalise independently
Signs the workspace is working
Pre-delivery review checklist
Evidence
- Major claims have source locations.
- Source status is visible.
- Limitations and contradictions remain visible.
- Missing evidence is identified.
Analysis
- Activities, outputs and outcomes are separated.
- Contribution is not confused with attribution.
- Assumptions and confidence are labelled.
- Conclusions match the strength of evidence.
Data, ethics and inclusion
- Calculations and periods are consistent.
- Identifiable information is protected.
- Consent and approved-use limits are respected.
- Country differences and missing voices are visible.
Communication and approval
- Language suits the audience.
- Claims are proportionate and understandable.
- The output is labelled as a draft.
- An authorised reviewer approves external use.
Practice exercise
Create a test workspace using non-sensitive programme material. Ask Claude to produce a one-page results summary, data-quality issues register, unsupported-claims list, reviewer questions and source-to-finding ledger.
Then deliberately add:
- one contradiction between two documents;
- one indicator with a missing denominator;
- one quotation that could identify a participant; and
- one cross-country comparison using incompatible periods.
Check whether Claude detects each problem, explains the risk and pauses for human verification instead of silently completing the task.
Copy-and-paste starter prompt
You are working inside an M&E and international development workspace. Before beginning: 1. Read context/MY-M&E-WORK.md. 2. Read system/M&E-SYSTEM-RULES.md. 3. Read the relevant project BRIEF.md. 4. Read SOURCE-REGISTER.md and EVIDENCE-LEDGER.md if present. 5. Review the required evidence, data dictionaries and templates. Use the workspace files as the main source of truth. Do not invent facts, values, findings, quotations, dates, approvals or causal explanations. Separate documented evidence, perception, calculation, interpretation, evaluative judgement, recommendation and unresolved question. Preserve claim-to-source traceability. Identify contradictions, missing evidence, incomparable periods and uncertainty. Protect personal, safeguarding and security-sensitive information. Preserve country and stakeholder differences. Pause when a stop condition in the system rules applies. Save generated work in the specified outputs folder. Create a reviewer-questions file and an evidence or assumption log. Treat every generated deliverable as a draft requiring human review.
Frequently asked questions
Do I need coding skills?
No. The workspace uses ordinary folders and text files. Markdown is useful because it is lightweight and easy to edit.
Do I need every file described here?
No. Start with the two reusable instruction files, a project brief and source register. Add an evidence ledger, data dictionary or specialised rules when the work requires them.
Can I use Word, Excel, CSV and PDF files?
Use the file types supported in your Cowork environment. Export data from systems such as KoboToolbox, ODK, spreadsheets or reporting platforms only when the export is authorised and necessary.
Should Claude always ask questions first?
No. It should ask when a missing detail, conflict or permission could materially change the result. Routine formatting does not require a long interview.
Can the workspace replace official records management?
No. It is a controlled working area for AI-assisted tasks. Approved records should remain in the organisation’s authorised systems.
Can Claude prepare a final evaluation report?
It can organise evidence, build matrices, test consistency and draft sections. Qualified professionals must review and approve findings, conclusions, ratings and recommendations.
Can teams share the same rules?
Yes, but the team should agree on permissions, source hierarchy, terminology, country handling, review ownership and which decisions remain human-only.
Final takeaway
Claude does not need access to every file. It needs the right evidence, explicit boundaries, a clear task and accountable human review.
Continue learning with EvalCommunity
Develop practical skills for applying AI responsibly across evaluation design, evidence synthesis, qualitative and quantitative analysis, reporting and M&E workflows.
Responsible-use note: Apply organisational policies, donor requirements, data-protection obligations, safeguarding procedures, local context and qualified professional judgement.
Suggested disclosure: “This draft was prepared with AI assistance and reviewed by the evaluation team of EvalCommunity.”
