20 useful ChatGPT prompts for social workers and programme teams
Copy these prompts when you need a first draft for a meeting, programme plan, volunteer message, training activity, public-source scan, report or plain-language edit. Each one keeps the task bounded and leaves judgement with your team.
In brief
A useful prompt names the task, audience, source material, format and limits. Give ChatGPT approved public, synthetic, anonymised or aggregated inputs, then have a person check the result before it is used.
Start with the job in front of you
The best ChatGPT prompts for social workers are usually ordinary work instructions. Say what you need, who will read it, which source material it may use, and what a good answer should look like. OpenAI's own prompt guidance recommends clear, specific instructions and iterative refinement.[1] If the first draft misses the mark, point to the part that needs changing instead of starting from scratch.
Replace every square-bracketed placeholder before sending a prompt. Delete any instruction you do not need. The prompts below ask for assumptions, gaps or checks because a smooth answer can still be wrong. They are drafting aids. A staff member remains responsible for source checking, professional judgement, safeguarding, approval and the final communication.
Meetings
Use these for the work around a meeting: shaping an agenda, turning safe notes into actions, and preparing questions. Keep a human record of what was actually agreed.
1. Build a focused meeting agenda
Create a practical agenda for a [LENGTH]-minute [MEETING TYPE] with [PARTICIPANTS OR ROLES]. The purpose is [PURPOSE]. By the end, we need to decide or produce [OUTCOMES]. Use these safe background points: [ANONYMISED, AGGREGATED, SYNTHETIC OR PUBLIC CONTEXT].
Return a table with time, agenda item, lead role, method and intended output. Reserve five minutes to confirm decisions, owners and deadlines. Flag any outcome that cannot realistically fit the time. Do not invent case details, decisions or participant views.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.2. Turn safe meeting notes into an action list
Organise the following meeting notes into a working record for [TEAM OR COMMITTEE]. The notes must be anonymised, non-confidential and approved for use in this tool: [PASTE SAFE NOTES].
Use four sections: decisions recorded, actions, open questions and items to verify. Put actions in a table with action, named role or placeholder owner, due date, dependency and status. Write 'not recorded' where the notes do not support an answer. Preserve uncertainty and disagreement. Do not turn suggestions into decisions or add commitments. End with three questions the meeting chair should resolve before circulating the record.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.3. Prepare constructive questions for a partner meeting
Draft [NUMBER] constructive questions for a meeting with [PARTNER TYPE] about [PUBLIC OR NON-CONFIDENTIAL TOPIC]. Our aim is [AIM]. Known facts from approved sources are: [FACTS]. Matters we need to understand are: [GAPS].
Group the questions under outcomes, roles, resources, risks and next steps. Make them open enough to invite useful detail and specific enough to avoid vague answers. Add a short note after each question explaining what it helps clarify. Avoid adversarial wording and do not assume the partner has agreed to anything.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.Programme planning
Planning prompts work better when they expose choices. Ask for assumptions, measures and unanswered questions alongside the draft plan. NCSS places digital work within wider organisational needs and readiness in its Social Services Digitalisation Playbook.[5] The same discipline helps with any programme change.
4. Sketch a programme logic model
Create a first-draft logic model for a [PROGRAMME TYPE] serving a broadly defined group: [NON-IDENTIFYING TARGET GROUP]. The need, based on [PUBLIC OR AGGREGATED SOURCE], is [NEED]. Available inputs are [PEOPLE, TIME, BUDGET RANGE, PARTNERS OR ASSETS]. Planned activities are [ACTIVITIES].
Return a table covering inputs, activities, outputs, short-term outcomes, longer-term outcomes, assumptions and external factors. Distinguish outputs from outcomes. Suggest no more than [NUMBER] measurable indicators, with a possible data source and collection burden for each. List claims that need evidence. Do not invent prevalence figures or promise impact.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.5. Stress-test a programme idea
Act as a constructive programme reviewer. Review this early idea: [DESCRIBE IDEA USING SYNTHETIC, AGGREGATED OR PUBLIC INFORMATION]. It aims to achieve [OUTCOME] for [BROAD GROUP] within [TIMEFRAME] and [RESOURCE LIMIT].
Identify the five most important assumptions. For each, explain what could go wrong, what evidence would reduce uncertainty, and one small reversible test. Then list possible access barriers, unintended burdens and delivery dependencies. Finish with a go, revise or pause recommendation for the test only, with reasons. Do not assess any individual, predict personal outcomes or invent evidence.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.6. Turn a goal into a 90-day work plan
Turn this team goal into a realistic 90-day work plan: [GOAL]. Context: [TEAM CAPACITY], [FIXED DATES], [BUDGET OR RESOURCE LIMITS], [DEPENDENCIES] and [KNOWN RISKS].
Break the plan into weeks 1-2, weeks 3-6, weeks 7-10 and weeks 11-13. For each phase, give deliverables, owner roles, decision points, evidence to collect and stop or change conditions. Include a short 'not doing now' list to protect scope. Mark every assumption and any missing information that could change the sequence. Keep the plan suitable for a team check-in, not a funding promise.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.Volunteer communications
A good volunteer message makes the requested action easy to find. Give the model the approved facts, channel and tone. Check dates, links, role expectations and contact routes yourself.
7. Draft a volunteer recruitment message
Draft a volunteer recruitment message for [CHANNEL] about [ROLE]. Use only these approved facts: [PURPOSE], [TASKS], [LOCATION OR REMOTE ARRANGEMENT], [DATES], [TIME COMMITMENT], [TRAINING], [ELIGIBILITY], [SUPPORT], [APPLICATION LINK] and [CLOSING DATE]. The audience is [AUDIENCE].
Write a clear headline and a body of no more than [WORD LIMIT] words. Lead with the practical contribution volunteers will make. State the commitment and selection process plainly. Avoid urgency tricks, saviour language, guaranteed outcomes and claims not supplied. End with one direct action. List any missing fact after the draft instead of guessing.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.8. Write a volunteer shift reminder
Write a warm, concise reminder for volunteers assigned to [ACTIVITY] on [DATE] at [TIME] at [APPROVED LOCATION]. Include [ARRIVAL INSTRUCTION], [WHAT TO BRING], [DRESS OR ACCESS REQUIREMENTS], [STAFF CONTACT ROUTE], [CANCELLATION PROCESS] and [WEATHER OR CONTINGENCY PLAN]. Channel: [EMAIL, WHATSAPP OR SMS]. Tone: [TONE].
Put the essential action and timing first. Use bullets if the channel allows. Include one sentence reminding volunteers to follow [CONFIDENTIALITY OR SAFEGUARDING POLICY] without describing any participant. Keep within [WORD OR CHARACTER LIMIT]. Flag missing logistics separately.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.9. Respond to volunteer feedback
Draft a reply to this [ANONYMISED, NON-CONFIDENTIAL] volunteer feedback: [PASTE FEEDBACK]. The reply will come from [ROLE] through [CHANNEL]. We can confirm these facts: [APPROVED FACTS]. We cannot yet confirm: [OPEN ITEMS].
Acknowledge the specific concern without becoming defensive. Separate what we heard, what we can do now, what needs review and when the volunteer can expect an update. Do not promise an outcome or admit fault unless the supplied facts support it and approval has been given. Keep the reply under [WORD LIMIT] words.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.Training
Training materials need a clear objective and a chance to practise. The model can help shape a session, scenario or quiz. A subject lead should still check accuracy and decide what participants are expected to do in real situations.
10. Build a short staff training outline
Design a [LENGTH]-minute training outline for [STAFF OR VOLUNTEER AUDIENCE] on [TOPIC]. By the end, participants should be able to [TWO OR THREE OBSERVABLE LEARNING OBJECTIVES]. Use these approved references: [POLICY, PUBLIC GUIDANCE OR TRAINING SOURCE]. Delivery mode: [IN PERSON OR ONLINE]. Group size: [NUMBER].
Create a timed sequence with opening, explanation, practice, debrief and check for understanding. Include facilitator notes, materials and one accessible alternative for participants who cannot join the main activity. Mark any content that requires a subject-matter check. Do not invent policy requirements.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.11. Create a safe practice scenario
Create one clearly fictional practice scenario for training on [SKILL OR POLICY]. Audience: [AUDIENCE]. Setting: [GENERIC SETTING]. The scenario should let participants practise [DECISION OR BEHAVIOUR] without requiring disclosure of personal experience.
Use invented names and neutral details. Provide: a 150-word scenario, three discussion questions, facilitator guidance, common weak responses and signs of a sound response. Avoid trauma detail, stereotypes, diagnosis and a single obvious 'perfect' answer. State that the scenario is synthetic. Link every teaching point to [APPROVED POLICY OR PUBLIC SOURCE] and flag anything the source does not settle.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.12. Draft a knowledge check
Create [NUMBER] knowledge-check questions from this approved training material: [PASTE NON-CONFIDENTIAL MATERIAL OR PUBLIC LINK TEXT]. Audience level: [BEGINNER, INTERMEDIATE OR REFRESHER]. Focus on [LEARNING OBJECTIVES].
Use a mix of multiple choice and short workplace situations. For each question, provide the correct answer, a brief explanation tied to the supplied material, and why each distractor is weaker. Avoid trick wording and avoid testing facts that are absent from the source. Include one 'find the answer in the policy' question. Finish with items a trainer should review for ambiguity before use.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.Public-source research
Research is useful only when another person can retrace it. Ask for links, dates, quotations and gaps. Open each cited page yourself. If browsing is unavailable, require the model to say so and work only from text you provide.
13. Scan official guidance on a topic
Research current official guidance on [TOPIC] for [JURISDICTION OR SECTOR]. Prioritise government, regulator and statutory-body sources published or updated after [DATE]. Exclude blogs, vendor pages and search-result snippets unless I ask for them.
Return a table with issuing body, document title, publication or update date, direct URL, relevant section and a short verbatim quotation. Then summarise what the sources clearly say, what remains unclear and what may have changed. Do not invent citations. If you cannot browse or open a page, say so and do not claim it was checked.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.14. Compare public programmes or service models
Compare [NUMBER] publicly documented programmes or service models addressing [NEED] in [PLACE OR SECTOR]. Use sources published by the programme operator, funder, government or an independent evaluator. Cut-off date: [DATE]. Comparison fields: [ELIGIBILITY, DELIVERY MODEL, STAFFING, COST, REACH, OUTCOMES OR OTHER FIELDS].
For every factual statement, give a direct source link and date. Separate reported facts from your interpretation. Use 'not found' where public evidence is missing. Do not rank the programmes unless I provide criteria. End with five questions our team should investigate before adapting any element to [OUR NON-CONFIDENTIAL CONTEXT].
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.15. Build a source-checking table
Check the following draft claims about [TOPIC] against current public sources: [PASTE CLAIMS WITHOUT CONFIDENTIAL INFORMATION]. Preferred source types: [OFFICIAL STATISTICS, REGULATOR, PRIMARY RESEARCH OR OTHER]. Geographic scope: [PLACE]. Evidence cut-off: [DATE].
Create a table with claim, verdict (supported, partly supported, unsupported or outdated), best direct source URL, source date, exact supporting passage, caveat and safer wording. Open the source rather than relying on a search snippet. If a claim cannot be verified, state what you searched and leave it unresolved. Do not fill gaps with plausible figures or citations.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.Reports
For reports, constrain the model to the evidence in front of it. Ask it to show missing data and calculation needs. Verify totals separately and preserve the distinction between observation, interpretation and recommendation.
16. Turn aggregate results into a report outline
Create a report outline from these approved aggregate results: [PASTE AGGREGATED DATA OR PUBLIC RESULTS]. Programme aim: [AIM]. Reporting period: [DATES]. Audience: [AUDIENCE]. Required headings: [HEADINGS].
Under each heading, list the evidence available, the claim it may support, the caveat, and any calculation or source check still needed. Separate outputs, outcomes and participant feedback. Do not infer causation from change over time or generalise beyond the data. Do not calculate figures unless all required values and definitions are present. Finish with a short list of questions for the programme lead before drafting prose.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.17. Draft a concise progress update
Draft a [WORD LIMIT]-word progress update for [FUNDER, BOARD, PARTNER OR INTERNAL TEAM] using only these approved facts: [MILESTONES], [AGGREGATE RESULTS], [SPEND OR RESOURCE POSITION], [CHALLENGES], [CHANGES] and [NEXT PERIOD ACTIONS]. Reporting period: [DATES]. Tone: [DIRECT, FORMAL OR CONVERSATIONAL].
Use the headings progress, evidence, issues, response and next steps. Preserve every qualifier in the source notes. If a required fact is missing, write TO CONFIRM instead of guessing. Avoid celebratory claims that go beyond the evidence. End with decisions or support needed from the reader, if supplied.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.18. Review a report for overclaiming
Review this draft report section for claims that outrun the evidence: [PASTE NON-CONFIDENTIAL DRAFT]. The available evidence is [AGGREGATED DATA, PUBLIC SOURCES OR APPROVED FINDINGS]. Intended audience: [AUDIENCE].
Make a table with original wording, concern, evidence available and proposed revision. Check especially for causal claims, unsupported superlatives, vague quantities, selective time periods, missing denominators and confusion between participation, satisfaction and outcomes. Preserve claims that are supported. Flag statements that need a source or subject-matter review. Do not create figures, quotations or participant stories.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.Plain-language editing
Editing is a strong low-consequence use because the team already owns the source text. Tell the model which facts, duties and terms must survive the rewrite. Compare the revision line by line before approval.
19. Rewrite a service notice in plain language
Rewrite this approved service notice for [AUDIENCE] in plain [LANGUAGE OR ENGLISH LEVEL]: [PASTE NON-CONFIDENTIAL NOTICE]. The reader needs to know [KEY ACTION]. Keep these terms, dates, eligibility conditions, contact routes and legal or policy wording unchanged: [MUST-KEEP DETAILS]. Desired length: [WORD LIMIT].
Lead with what the reader needs to do. Use short sentences, familiar words and useful headings or bullets. Explain necessary technical terms once. Do not add reassurance, rights, exceptions or service promises absent from the original. After the rewrite, list any sentence whose meaning may have shifted and any detail that still seems hard to understand.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.20. Make a form or survey easier to answer
Review these [FORM OR SURVEY] questions for clarity and burden: [PASTE QUESTIONS WITHOUT PERSONAL RESPONSES]. Purpose: [PURPOSE]. Audience: [AUDIENCE]. Reading level or language need: [NEED]. We must collect: [REQUIRED INFORMATION]. We may remove or make optional: [FLEXIBLE INFORMATION].
For each item, identify confusing wording, double questions, assumptions, unnecessary sensitivity and response options that may not fit. Propose a plain-language revision and explain why. Suggest 'prefer not to say' or 'not applicable' only where appropriate. Do not decide legal necessity or invent consent wording. List items needing privacy, accessibility or subject-matter review.
Data boundary: Do not include, request, infer or reproduce client-identifying data. Use only synthetic, anonymised, aggregated or approved public information.Get a better second draft
Read the first answer with a pen in hand. Ask: Which statement lacks a source? What did the model assume? What would confuse the intended reader? Which sentence sounds more certain than the evidence? Then give one focused revision instruction, such as shorten the opening to 60 words and keep every date, or turn the unresolved points into questions for the programme lead.
If your team wants to move beyond individual experiments, use the responsible AI implementation guide for social service agencies. For a bounded first test, read why your first AI project should probably be boring.
Sources
- [1] OpenAI Help Center, Prompt engineering best practices for ChatGPT
- [2] Personal Data Protection Commission Singapore, Advisory Guidelines on Use of Personal Data in Generative AI
- [3] OpenAI Help Center, Data Controls FAQ
- [4] OpenAI, Usage policies
- [5] National Council of Social Service, Social Services Digitalisation Playbook
About the author
Darren writes for Social Tech Guild about practical uses of technology in Singapore's social service sector.
Choose one safe task to try
Bring a routine drafting task, an approved source and your review criteria. We can help your team shape a small, accountable AI experiment around the work you already do.
Discuss a practical AI pilotCould a small tool make your team’s work lighter?
Tell me about a task that keeps taking time. We can look at it together and see whether a small volunteer project could help.
Please do not include client-identifying information.