Claude Duet for Evaluators and Development Professionals
EvalCommunity Academy · AI Tools for M&E Practice
Claude Duet for Evaluators and Development Professionals: A Practical Tutorial
What Anthropic’s proactive AI agent can do, how to set it up, how to build it into your daily M&E work, and how to use it responsibly with programme and evaluation data.
Level: Beginner to intermediate · Reading time: about 35 minutes · Updated: November 2026
Availability note (read first)
Claude Duet opened as a research preview on 21 October 2026. Access is through a waitlist for Claude Pro and Max subscribers with US accounts, on iOS, Android and the web (claude.ai). Team and Enterprise plans are not yet included, and Microsoft 365 is not yet supported. Many EvalCommunity members work outside the US or in Microsoft 365 organisations, so treat this tutorial as preparation: understand the tool, plan your use cases, and be ready when access widens.
Learning objectives
By the end of this tutorial you will be able to:
- Explain what a proactive AI agent is and how Claude Duet differs from a regular chatbot.
- Describe Duet’s core capabilities and current limits.
- Set up Duet step by step and build a daily and weekly routine for your role.
- Write clear rules and requests that produce useful results.
- Identify realistic use cases across the evaluation and project cycle.
- Apply data protection, ethics and quality safeguards before delegating M&E work to an agent.
1. What is Claude Duet?
Most of us use AI as a chatbot: you ask, it answers, and nothing happens until you ask again. Claude Duet is a proactive agent: a separate app in which Claude works in the background across the accounts you connect (email, calendar, cloud storage and other tools), flags what needs your attention, and carries out tasks you assign, even when your laptop and phone are off.
Think of it as a junior programme officer who reads your inbox every few hours, keeps a running list of open tasks, prepares drafts, and comes back to you for sign-off before anything leaves your name.
Regular AI chat
- You start the work, every time
- Runs in your open chat window
- Sees only what you paste or upload
- Gives text answers
- Takes no actions
Claude Duet (agent)
- You or Duet start the work; it flags things itself
- Runs on its own cloud computer and browser
- Uses connected apps and websites it opens
- Produces docs, sheets, dashboards, drafts
- Prepares actions, then waits for your approval
2. How it works behind the scenes
Four design choices shape what you can rely on Duet for.
Its own cloud computer
It runs on a virtual machine with its own browser, so it can work through sites with no connector, such as donor portals or statistics office websites.
Checks every few hours
It reviews accounts periodically, not continuously, to protect your usage limits. Keep your own reminders for hard deadlines.
One conversation, organised
You message it in plain language, typed or spoken. Recurring topics can live in dedicated channels that share the same memory.
You set the interruption rules
You decide what deserves an immediate message. Everything else waits for a daily morning brief.
3. Core capabilities, translated for M&E work
Vendor demos use bakeries and online shops. Here is what each capability means in evaluation and programme work.
4. How to set up and use Duet in your daily work
Part A is a first-day setup of about 45 minutes. Part B is a daily and weekly routine. Part C gives playbooks for three roles, Part D a library of standing tasks, and Part E fixes for common problems.
Part A. Your first day: setup in seven steps
Get access
Join the waitlist from the Claude Duet product page with the email of your Claude Pro or Max account. Invitations go out in waves.
Introduce yourself and your work (5 min)
Duet’s judgement depends on context. Tell it who you are, what you work on and how you work.
I’m an independent evaluator specialising in education and gender equality. I have two active assignments: a mid-term evaluation of a girls’ education project in Kenya (client: an international NGO) and a final evaluation for a UN agency in Jordan. I work 8:00–17:00 Eastern time, Monday to Friday. My priorities are client deadlines, data collection progress and new evaluation ToRs.
Connect apps one at a time (10 min)
Connect email, then calendar, then cloud storage. After each one, say what it is for and what to ignore. Start small: these three are enough for the first month.
This Gmail account is for client work only. Ignore newsletters, LinkedIn notifications and promotions. Anything from the Kenya project team or the Jordan evaluation manager is high priority.
Point it to your key working files (5 min)
Name the folders with your workplans and tracking sheets. Never connect folders holding identifiable respondent data (see Section 7).
Workplans and deliverable schedules are in Drive under “Evaluations / [project] / Management”. Use them as the reference for all deadlines. Do not open the “Raw data” folders.
Set interruption and approval rules (10 min)
Define what is urgent, when your brief arrives, and what must never happen without you.
Urgent (message me now): client emails, deadlines within 72 hours, field team problems. Morning brief: 7:30 my time on weekdays, five items maximum, most important first. Always draft, never send: anything to clients, donors, officials or partners.
Create your channels (5 min)
One channel per assignment (e.g. “Kenya MTE”), plus Business development (ToRs, tenders, proposals), Admin and finance (invoices, contracts, travel) and Learning (guidance, webinars, methods).
Run a low-risk test task (10 min)
Pick something easy to check, compare the output with what you know, and correct Duet in the conversation.
List every deliverable and deadline for my two assignments for the next 60 days, using the workplans in Drive. Put them in a sheet sorted by date and tell me which ones conflict.
Then pilot for two weeks. Treat Duet like any new tool in your M&E system: low-risk tasks first, review every output, log errors, then decide what to scale. Background checks and briefs count against your plan’s usage limits, so watch consumption.
Part B. Your daily and weekly routine
Consistent habits, not one-off requests, are where an agent pays off. This routine takes 20 to 30 minutes a day.
Part C. Role playbooks
Independent evaluator or consultant
Connect: work email, calendar, cloud storage with assignment folders.
Hand over: ToR tracking, interview scheduling, deadline tracking, invoice reminders, annex compilation, first drafts of background sections.
Keep: methodology, evaluative judgements, sign-off on client communication, pricing.
For each active assignment, track deliverables against the workplan. Warn me 7 days and 2 days before each deadline with a short status note and what is still missing.
MEAL officer or M&E manager in an NGO
Connect: work email and calendar (if your organisation permits it and is not on Microsoft 365), shared folders with aggregated indicator data.
Hand over: chasing partner reports, collating submissions, checking tracking tables for blanks or outliers, dashboard updates, donor report drafts.
Keep: data quality decisions, interpretation, complaints handling, anything touching beneficiary records.
On the 3rd of each month, check which partners have submitted their report to the shared folder and draft reminders to the rest for my approval. On the 6th, list any ITT values that are blank, over 150% of target, or changed more than 30% from last month.
Programme or project manager
Connect: email, calendar, project folders, budget tracking sheets (aggregated, not payroll records).
Hand over: meeting prep packs, action-item follow-up, workplan slippage alerts, burn-rate summaries, reporting calendars.
Keep: staff matters, partner relationships, budget reallocation, donor negotiations.
The day before each partner or donor meeting on my calendar, prepare a one-page brief: purpose, open action items from last time, relevant figures from the tracking sheets, and three questions I might be asked.
Part D. Standing tasks library (copy and adapt)
Standing tasks repeat without you asking again. Start with two or three, since each uses part of your plan’s limits.
Weekly ToR scan
Every Monday, check [job and tender sites] for evaluation ToRs in [sectors] and [regions]. Add matches to my “Opportunities” sheet with deadline, donor, budget and expertise. Summarise the best three in my brief.
Fieldwork progress digest
During fieldwork, read the team’s daily update emails and send me one evening summary: interviews completed vs planned, issues raised, decisions needed.
Reporting calendar
Keep a calendar of donor reporting deadlines from the grant agreements in [folder]. Remind me 21, 10 and 3 days before each, with a checklist of missing inputs.
Action-item follow-up
Track action items from meeting notes I share. If someone else’s item is 3 days overdue, draft a friendly follow-up for my approval.
Learning digest
Every Friday, check [networks and publications I follow] for new guidance on [topics]. Send a short list with one-line summaries and links.
Part E. Fine-tuning when things go wrong
5. Writing good rules and requests
Duet needs no special syntax, but clear instructions improve its judgement. Good requests name the outcome, the source, the format and the review point.
WEAK: “Do my donor report.”
STRONG: “Using the Q3 tab of the Indicator Tracking Table in the ‘Project X / MEAL’ Drive folder, draft the results section of the quarterly donor report. Compare actuals to targets, flag any indicator below 80% of target with a one-line explanation from the field notes, and give me a Google Doc to review. Don’t send anything.”
WEAK: “Find me evaluation jobs.”
STRONG: “Every Monday, check the tender and job sites I listed for final evaluation ToRs in WASH or livelihoods, budget over USD 30,000, deadline at least 10 days away. Put them in a sheet with link, donor, deadline and required expertise, and summarise the top three in my brief.”
WEAK: “Schedule the interviews.”
STRONG: “Draft invitations to the 12 stakeholders in the KII list sheet, offering 45-minute slots from my calendar between 10 and 20 November. Show me the drafts before sending, and log confirmed times in the sheet.”
6. Use cases across the evaluation and project cycle
Good fit tasks are repetitive, low-risk and easy to verify. Use with care tasks need closer review. Keep human-led tasks rely on judgement, relationships or ethical responsibility.
Business development
Good fit: tracking ToRs and tenders; deadline calendars; collecting past CVs and references
With care: first drafts of technical proposals and methodology sections
Human-led: pricing, go/no-go decisions, partner negotiations
Inception
Good fit: organising project documents; building a document review log
With care: draft evaluation matrix and theory of change narrative
Human-led: final evaluation questions, sampling and ethics decisions
Fieldwork
Good fit: interview scheduling; logistics checklists; daily progress summaries
With care: monitoring data collection progress against targets
Human-led: anything involving respondents’ personal data, consent or safeguarding
Analysis
Good fit: descriptive summaries of anonymised, aggregated indicator data
With care: spotting anomalies and proposing hypotheses to test
Human-led: causal claims, contribution analysis, evaluative judgements
Reporting
Good fit: formatting, annexes, reference checks, validation workshop decks
With care: executive summary drafts; quarterly donor report drafts
Human-led: findings, conclusions, recommendations, management responses
Programme MEAL
Good fit: reporting-cycle reminders; collating partner submissions; dashboard updates
With care: learning brief drafts; summarising feedback trends
Human-led: responding to complaints and feedback from affected people
Worked example: a week with Duet during a mid-term evaluation
Monday brief: “Two items need you. The programme manager sent revised Year 2 beneficiary numbers that differ from the inception report; the differences are in a sheet. Four of 12 KII invitations are unanswered; follow-up drafts are ready.”
You: “Send the follow-ups. Flag the beneficiary numbers again on Wednesday if the programme manager hasn’t explained the change.”
Duet: shows each follow-up for approval. You approve three and rewrite the one to a senior government official yourself.
Thursday: “Draft the document review section of the inception report from the ‘Desk review’ folder, with a source for every claim.” You check every reference against the originals before it goes near the client.
7. Responsible use: data protection, ethics and quality
Agents raise the stakes because they read your accounts and act across them, and your obligations to respondents, clients and donors do not transfer to a tool. Anthropic’s safeguards help: Duet does not see your passwords or access tokens directly, consequential actions need approval, and there are protections against prompt injection (hidden instructions planted in emails or web pages). They do not replace your own controls.
Do not delegate
- Personal or identifiable data of participants, respondents or interviewees.
- Safeguarding, protection or PSEA case information.
- Data from contexts where exposure could put people at risk.
- Client data your contract, data-sharing agreement or ethics approval does not allow you to process with third-party AI services.
Responsible-use checklist before you start
- ☐ I have checked my client contracts and my organisation’s AI policy.
- ☐ Connected accounts contain no identifiable respondent data.
- ☐ Nothing goes to clients, partners or donors without my approval.
- ☐ I verify every figure, quote and reference in AI-produced drafts.
- ☐ I disclose AI assistance where the client or evaluation policy requires it.
- ☐ I keep my own reminders for time-critical deadlines.
- ☐ I know that disconnecting an app stops further access, but Duet keeps what it already learned.
8. Current limitations
9. Knowledge check
Tap each question to reveal the answer.
1. What makes Claude Duet “proactive”?
It works in the background across connected accounts, flags what needs attention and carries out assigned tasks without waiting for you to open a chat.
2. Why keep your own reminders for a proposal deadline?
Duet checks accounts every few hours rather than continuously, and as an early version it can misjudge urgency.
3. Which four kinds of action always need your approval?
Sending, spending, signing and deleting.
4. A cleaned survey dataset still includes respondents’ names and villages. Should you connect its folder?
No. Identifiable respondent data stays out of agent-connected accounts. Use anonymised, aggregated data only, and only where your contract and ethics approval allow it.
5. What should you do in your first two weeks with Duet?
Run a pilot: low-risk tasks only, review every output, log errors, adjust your rules each Friday, then decide what to scale.
6. Name one evaluation task that should stay human-led, and why.
Findings, conclusions and recommendations; causal or contribution claims; ethics and sampling decisions. They require professional judgement and accountability an evaluator cannot delegate.
10. Key takeaways
- Claude Duet is a background agent, not a chatbot: it monitors, prepares and acts, with your approval for anything consequential.
- Set it up deliberately: introduce your work, connect a few apps, set clear rules, then pilot for two weeks.
- Its best early uses in M&E are coordination work: tracking, scheduling, collating and drafting.
- Evaluative judgement, ethics and relationships with respondents and clients stay with you.
- Access is limited for now (US, Pro/Max, waitlist, no Microsoft 365), so use this time to plan pilots and update your AI policies.
Build your AI skills for evaluation
Explore more EvalCommunity Academy courses and tools on responsible AI in monitoring, evaluation, accountability and learning.
This tutorial reflects publicly reported information about the Claude Duet research preview as of November 2026. Features, availability and limits may change; check Anthropic’s official announcements before making organisational decisions. EvalCommunity is not affiliated with Anthropic.
