Master the art of communicating with AI — transform vague requests into powerful, reliable prompts for healthcare administration workflows.
⏱ 60 minutes · 5 exercises
The 4 Pillars of Great Prompts
Every effective prompt on AgentSea follows these four principles. Master them and your agents will produce consistent, reliable results every time.
🎯
1. Clear
State exactly what you want. No ambiguity. "Summarize" is vague — "Extract the top 5 action items with owners and deadlines" is clear.
📐
2. Specific
Define the format, length, and structure. "Give me a table with columns: Item, Owner, Deadline, Status" leaves no room for guessing.
🏥
3. Contextual
Tell the AI who it is and who it's writing for. "You are a Senior Procurement Analyst writing for the Finance Director" sets the right tone.
🛡️
4. Constrained
Set boundaries. "Only use data from the uploaded document. Never fabricate figures. Flag missing information as NOT PROVIDED."
🔥 Warm-Up: Spot the Problem
Look at these prompts that a hospital admin team might use on AgentSea. What's wrong with each one? Discuss with your team (2 minutes).
❌ Prompt A
"Summarize this document"
Problem:No format specified, no audience, no length limit — you'll get a different result every time
❌ Prompt B
"Tell me everything about this vendor proposal and whether we should go with them"
Problem:"Everything" is unbounded, asking for a decision without criteria, no guardrails on what "should" means
❌ Prompt C
"Write me an email to the team about the budget"
Problem:No tone, no key points to include, no audience context, no purpose (inform? request? escalate?)
💡 Click the blurred text to reveal the answer. Discuss with your team first!
Exercise 1: Transform a Vague Prompt
Take the vague prompt below and rewrite it using all 4 pillars. Try it on AgentSea with a sample document and compare the results.
Scenario: Hospital Procurement
You've received a vendor proposal for medical equipment maintenance services. You need to assess it before the next procurement committee meeting.
❌ BEFORE — Vague prompt
Look at this vendor proposal and tell me what you think.
🤔 Your turn: Rewrite this prompt using the 4 pillars. Consider:
Clear: What exactly do you want extracted?
Specific: What format should the output be in?
Contextual: Who are you? Who is the audience?
Constrained: What should the AI NOT do?
👀 Reveal: Example strong prompt
✅ AFTER — Powerful prompt
You are a Senior Procurement Analyst with 10 years of experience in healthcare vendor assessment.
Read the attached vendor proposal and produce:
1. **Summary Table** with columns:
| Vendor Name | Service Offered | Contract Duration | Annual Cost (SGD) | SLA Uptime % | Key Risks |
2. **Compliance Checklist** — check these against our requirements:
- [ ] ISO 13485 certified (medical devices)
- [ ] PDPA compliant data handling
- [ ] 24/7 emergency response capability
- [ ] Local service team in Singapore
- [ ] References from healthcare sector
3. **Recommendation:** PROCEED / REVIEW FURTHER / REJECT
- One paragraph justification (max 100 words)
- List any missing information that would change the recommendation
Rules:
- Only use information from the attached document
- If any checklist item is not mentioned, mark as "NOT PROVIDED — verify with vendor"
- All costs in SGD
- Do not recommend PROCEED if any compliance item is missing
- Flag any cost that seems unusually low or high compared to typical healthcare maintenance contracts
✅ Try it: Open AgentSea → select Document Analysis skill → upload a sample document → paste your improved prompt. Compare the output quality with the vague version.
Exercise 2: The Persona Technique
The persona formula gives AI a "character" to play — making responses more consistent, expert-level, and appropriately toned for your audience.
The Formula
Title + Experience + Specialty + Characteristic + Behavior
"Senior Finance Manager with 12 years in healthcare budgeting, detail-oriented and conservative, who always cites source data"
Scenario: Write personas for these roles
For each scenario below, write a persona using the formula. Then test it on AgentSea — notice how the tone and depth of response changes.
Scenario
Your Persona (fill in)
A. Reviewing staff leave applications for HR
Try: "HR Operations Specialist with 8 years in healthcare workforce management..."
B. Drafting a memo to department heads about budget cuts
Try: "Chief Financial Officer with 15 years in public healthcare finance..."
C. Summarizing patient feedback surveys (non-clinical)
Try: "Patient Experience Manager with 10 years in hospital quality improvement..."
D. Checking a new MOH circular against current policies
Try: "Compliance Officer with 12 years in healthcare regulatory affairs..."
EXAMPLE — Persona in action on AgentSea
You are a Senior Administrative Officer with 10 years of experience in hospital operations management. You are methodical, detail-oriented, and always prioritize patient safety considerations even in administrative decisions. You write in a professional but accessible tone suitable for department heads who are clinicians, not administrators.
Read the attached meeting minutes and produce:
1. A bullet-point summary of key decisions (max 5 bullets)
2. An action item table: | Action | Owner | Deadline | Priority (High/Medium/Low) |
3. Any items that may impact patient services — flag these with ⚠️
Only extract information explicitly stated in the document. If a deadline is not mentioned, write "TBD — follow up required".
💡 Why personas work: Without a persona, AI gives generic responses. With a persona, it adopts the expertise level, vocabulary, and judgment of that role. A "Senior Finance Manager" will flag budget risks that a generic AI wouldn't notice. A "Compliance Officer" will catch regulatory gaps that others miss.
Exercise 3: Structured Output
The biggest difference between a useful AI response and a useless one? Structure. Tell the AI exactly what format you want — tables, checklists, numbered sections — and it will deliver consistently.
Scenario: Monthly Department Report
You need to generate a monthly operations report from raw data. The report goes to the Hospital Director. Compare these two approaches:
❌ UNSTRUCTURED — What you get is unpredictable
Create a monthly report from this spreadsheet data.
✅ STRUCTURED — Consistent, professional output every time
You are a Hospital Operations Analyst preparing the monthly report for the Hospital Director.
From the attached spreadsheet, generate a report with EXACTLY this structure:
## Executive Summary (3 sentences max)
[Overall performance statement, key highlight, key concern]
## KPI Dashboard
| Metric | This Month | Last Month | Target | Status |
|--------|-----------|------------|--------|--------|
[Fill from data — use ✅ for on-target, ⚠️ for within 10%, 🔴 for missed by >10%]
## Highlights (max 3 bullets)
- [Positive achievement with specific number]
- [Positive achievement with specific number]
- [Positive achievement with specific number]
## Concerns Requiring Attention (max 3 bullets)
- [Issue + impact + suggested action]
- [Issue + impact + suggested action]
- [Issue + impact + suggested action]
## Next Month Focus Areas
1. [Priority action]
2. [Priority action]
3. [Priority action]
Rules:
- All percentages to 1 decimal place
- All costs in SGD with thousands separator
- If data is missing for any metric, write "Data not available — follow up with [department]"
- Keep total report under 400 words
- Tone: Professional, factual, no jargon
✅ Try it: On AgentSea, upload any spreadsheet with numbers (even a simple one you create). Test both prompts and compare the outputs. Which one would you confidently send to your director?
More Structure Patterns for Healthcare Admin
Use Case
Best Output Format
Why
Vendor comparison
Comparison table + recommendation
Side-by-side makes differences obvious
Policy compliance
Checklist with ✅/❌/⚠️ status
Quick visual scan for gaps
Meeting minutes
Decisions + Action items table
Actionable, easy to track
Budget analysis
Variance table + narrative
Numbers need context
Email drafting
Subject line + body + sign-off
Ready to send format
Risk assessment
Risk register (Impact × Likelihood matrix)
Standard risk format
Exercise 4: Guardrails & Negative Constraints
In healthcare, what AI must NOT do is often more important than what it should do. Guardrails prevent the AI from overstepping boundaries, fabricating data, or making decisions it shouldn't make.
The Three Types of Guardrails
🚫
Must NEVER
Hard boundaries the AI cannot cross. "Never provide medical advice", "Never fabricate data", "Never approve amounts above $50K"
⚠️
Escalate When
Conditions that trigger human handoff. "If confidence is low", "If amount exceeds threshold", "If patient data is involved"
📌
Data Grounding
Rules about information sources. "Only use uploaded documents", "Cite page numbers", "Say NOT PROVIDED for missing data"
Scenario: Write Guardrails for These Agents
For each agent below, write at least 3 "Must NEVER" rules and 2 "Escalate when" triggers. Think about what could go wrong in a healthcare context.
🏥 Agent A: Patient Feedback Summarizer
Reads patient satisfaction surveys and generates a summary report for the quality improvement team.
Must NEVER:
Include any patient names, NRIC, or identifiable information in the summary
Make clinical judgments about care quality
[Your turn — add more]
Escalate when:
Feedback mentions patient safety concerns or near-misses
[Your turn — add more]
💰 Agent B: Purchase Order Validator
Checks purchase orders against approved budget, vendor list, and procurement policy before routing for approval.
Must NEVER:
[Your turn — think about financial risks]
Escalate when:
[Your turn — what thresholds matter?]
📧 Agent C: Staff Communication Drafter
Drafts internal emails and memos for department heads based on bullet-point instructions.
Must NEVER:
[Your turn — what could go wrong with automated emails?]
Escalate when:
[Your turn — when should a human review before sending?]
👀 Reveal: Example guardrails for Agent B (Purchase Order Validator)
GUARDRAILS — Purchase Order Validator
## You Must NEVER
- Approve or recommend approval of any purchase order (you only validate — humans approve)
- Process POs from vendors not on the approved vendor list
- Accept POs without a valid budget code and cost center
- Override procurement policy limits regardless of justification given
- Disclose budget remaining balances to requestors (only to finance team)
## Escalate to Human When
- PO amount exceeds $25,000 SGD (requires Director approval)
- Vendor is new (not on approved list) — route to Vendor Management
- Budget code shows insufficient remaining balance
- PO is for controlled items (pharmaceuticals, medical devices, IT equipment)
- Same vendor has 3+ POs in the same week (potential split-ordering to avoid thresholds)
- Requestor and approver are the same person
## Data Grounding Rules
- Only validate against the uploaded approved vendor list and budget file
- If a vendor is not found in the list, state "VENDOR NOT ON APPROVED LIST — requires Vendor Management review"
- If budget data is unavailable, state "BUDGET VERIFICATION PENDING — cannot validate without current balance"
Exercise 5: Build a Complete Prompt
Now put it all together. Choose ONE of the scenarios below and write a complete prompt that combines all 4 pillars: persona, clear instructions, structured output, and guardrails. Test it on AgentSea.
Choose Your Scenario
📋 Scenario A: Audit Preparation
You have 20 policy documents that need to be checked against a new MOH audit checklist. The audit is in 2 weeks. Write a prompt that helps you systematically check each document.
Skills: Document Analysis + Document Generation
📧 Scenario B: Stakeholder Update
You need to send weekly project updates to 5 different stakeholder groups, each needing different levels of detail. Write a prompt that generates all 5 versions from one status report.
Skills: Document Analysis + Document Generation + Outlook
💰 Scenario C: Invoice Verification
Your team processes 50+ vendor invoices per week. Each needs to be checked against the PO, delivery receipt, and contract terms. Write a prompt for a 3-way match verification agent.
Skills: Document Analysis
🏗️ Scenario D: Onboarding Checklist
New staff onboarding requires 15+ documents to be prepared, customized with the new hire's details. Write a prompt that generates the complete onboarding pack from a single intake form.
Skills: Document Analysis + Document Generation
Your Complete Prompt Should Include:
✅ Persona — Title + Experience + Specialty + Characteristic + Behavior
✅ Purpose — One clear sentence of what the agent does
✅ Steps — Numbered workflow (what to read → what to do → what to output)
✅ Output format — Exact structure (tables, sections, bullet points)
✅ Rules — At least 3 positive rules ("must always...")
✅ Must NEVER — At least 3 negative constraints
✅ Escalation — At least 2 triggers for human handoff
✅ Data grounding — "Only use uploaded documents, cite sources"
✅ Test your prompt on AgentSea:
Does it produce the exact format you specified?
Try giving it incomplete data — does it say "NOT PROVIDED"?
Try asking something outside scope — does it refuse or escalate?
Is the tone right for your audience?
Would you trust this output enough to forward to your director?
📋 Prompt Engineering Cheat Sheet
Keep this reference handy when writing prompts for AgentSea. Print it or bookmark this page.
The Prompt Structure Template
UNIVERSAL PROMPT TEMPLATE FOR AGENTSEA
You are [PERSONA — Title + Experience + Specialty + Characteristic].
Your purpose is to [ONE SENTENCE — what this agent does].
## What You Do
When the user uploads [INPUT TYPE], you will:
1. [Read/Extract — what to pull from the document]
2. [Analyze/Compare — what to do with the information]
3. [Generate/Output — what to produce]
## Output Format
[EXACT structure — tables, sections, bullet points, word limits]
## Rules You MUST Follow
- Only use information from the uploaded document(s)
- If information is missing, state "NOT PROVIDED"
- [Domain-specific rule]
- [Tone/audience rule]
- [Format rule]
## You Must NEVER
- Fabricate or assume data not in the document
- [Domain-specific prohibition]
- [Safety/privacy prohibition]
- Provide medical, legal, or financial advice
## Escalate to Human When
- [Threshold trigger]
- [Confidence trigger]
- You are unsure about the correct interpretation
Quick Reference: Power Phrases
Instead of...
Say this...
Why it works
"Summarize this"
"Extract the top 5 key points as bullet points, max 20 words each"
Specific format + length constraint
"Is this good?"
"Evaluate against these 5 criteria and rate each 1-5 with justification"
Measurable, structured assessment
"Write an email"
"Draft a professional email to [audience] informing them of [topic]. Tone: [formal/friendly]. Max 150 words."
Audience + purpose + tone + length
"Check this document"
"Compare this document against [reference] and produce a compliance checklist with ✅/❌/⚠️ for each requirement"
Clear comparison + visual output
"Help me with this"
"You are a [persona]. Read the attached [document type] and produce [specific output format]"
Role + input + output defined
"What do you think?"
"Based only on the data provided, list 3 strengths and 3 risks. For each risk, suggest one mitigation action."
Grounded + structured + actionable
Healthcare-Specific Guardrails (Always Include)
🛡️ Standard Guardrails for Healthcare Admin Agents
Never include patient names, NRIC numbers, or identifiable health information in outputs
Never provide clinical advice, diagnoses, or treatment recommendations
Never approve expenditures — only validate and recommend (humans approve)
Always cite the source document and page/section for any claim
State "NOT PROVIDED" for any information not found in uploaded documents
Escalate to human when patient safety may be impacted
All monetary values in SGD
Follow PDPA guidelines — no personal data in summaries unless explicitly required
Common Mistakes to Avoid
❌ Don't
Ask open-ended questions ("what do you think?")
Forget to specify output format
Skip negative constraints
Use jargon the AI might misinterpret
Assume AI knows your context
✅ Do
Give exact format (table columns, bullet count)
Set word/length limits
Include "Must NEVER" rules
Define escalation triggers
Test with edge cases (missing data, out-of-scope requests)