A guided build everyone completes together. Give an agent one document, make it answer only from that document, and turn the answer into a brand-compliant briefing.
⏱ 60 minutes · guided build · no coding
What you'll build
You'll build the Guideline Assistant — an agent that reads one internal staff guideline and helps you understand it. It's the shared warm-up before you design an agent for your own work in F4. Everyone builds the same thing, so we move together.
This module teaches the single most important idea in the whole day: grounding. A general chatbot answers from everything it has ever read. A grounded agent answers from your document — and tells you honestly when the answer isn't there.
You're done when you can show all five
1
A system prompt with a role, the task, grounding rules, and escalation behaviour.
2
Document Analysis selected as a skill so the agent can read the guideline.
3
A covered answer that cites the relevant guideline section.
4
An uncovered answer that says "not covered" instead of inventing one.
5
A plain-language, brand-compliant briefing produced with Document Builder.
Step 3 — Upload the guideline, ask a covered question
48–56 min
Step 3 continued — ask the deliberately uncovered question
56–60 min
Step 4 + debrief — generate the briefing, then reflect
🧪 One shared build. We use one synthetic guideline and one system prompt so a room of 50 can keep pace. The deeper build — with full verification and guardrails — is the optional Use Case Deep Dive.
Before you start
Log in to AgentSea and go to My Agents. You'll create your own private agent here — nothing you build is shared until you choose to share it.
Have the guideline ready. Download the synthetic Staff Leave & Time-Off Guideline below, or use the copy your facilitator shares. You'll upload it in Step 3.
Keep your F2 system prompt handy. You drafted one in the previous module — we'll turn it into this agent's operating instructions.
This document is classifiedRestricted, Sensitive (Normal)— the level AgentSea is approved to handle. Always verify responses.
⚠️ Data privacy. Everything in this exercise is synthetic. Do not upload real staff records, patient information, or anything above Restricted, Sensitive (Normal). Follow PDPA at all times.
Step 1 · Configure the agent (≈10 min)
Create a new agent, then open the Configure tab. Create writes a setup for you from a plain-language description; Configure is where you take control of the exact instructions — which is what we want here, because the wording is the behaviour.
Paste the system prompt below into the agent's instructions. It's the F2 pattern — role · task · grounding rules · escalation — already written for the Guideline Assistant. Read it as you paste; every line does a job.
SYSTEM PROMPT — paste into AgentSea Configure
You are the Guideline Assistant — a careful, plain-speaking policy guide for Synapxe staff. You help colleagues understand an internal staff guideline. You are precise, patient, and you never guess.
## What you do
When the user uploads a guideline document, you will:
1. Answer their question using ONLY the content of that document.
2. Point to where the answer comes from — quote or name the section number and heading (for example, "Section 3.1 — Annual Leave").
3. Keep answers short and in plain language a non-specialist can act on.
## Rules you MUST follow
- Use only the uploaded guideline. Do not use outside knowledge or assumptions.
- Every answer must cite the section it came from.
- If the guideline does not cover the question, reply exactly: "This is not covered in the guideline." Then say which team to ask (see escalation).
- If the guideline is unclear or partly relevant, say so plainly rather than filling the gap.
- All figures, dates, and day counts must match the document exactly. Never round or estimate.
## You must NEVER
- Invent a policy, number, or entitlement that is not in the document.
- Give legal, medical, or financial advice.
- Reveal or infer personal data about any individual.
## When to escalate
If the answer is not in the guideline, or the user asks for an exception, respond:
"This is not covered in the guideline. Please check with People & Culture (HR), or your Head of Department for exceptions (see Section 8)."
## Tone
Warm, direct, and brief. You are a helpful colleague, not a legal notice. Remind the user to verify anything important before acting on it.
✅ Checkpoint: Your agent has a clear role, a defined task, an explicit grounding rule ("use only the uploaded guideline"), a citation requirement, and an escalation path. That's success criterion #1.
Why this wording matters: the line "If the guideline does not cover the question, reply exactly: 'This is not covered in the guideline'" is what stops the agent from making things up. You'll test it directly in Step 3.
Step 2 · Add the skill (≈4 min)
A system prompt tells the agent how to behave. A skill gives it something it can do. To read an uploaded document, the agent needs the Document Analysis skill.
In the Configure tab, open Skills and enable Document Analysis. Leave the others off for now — we want the simplest agent that works.
Skill available today
What it lets the agent do
Use it here?
Document Analysis
Read and reason over an uploaded PDF, Word, or Excel file
✅ Yes — this is how it reads the guideline
Document Builder
Produce a formatted, brand-compliant document
Later, in Step 4
Microsoft Exchange (Outlook)
Read and draft email and calendar items
Not needed today
✅ Checkpoint: Document Analysis is switched on. That's success criterion #2.
Looking ahead: AgentSea's skill set is growing — connectors to document libraries and knowledge spaces, and reusable custom skills, are on the roadmap as concepts subject to platform readiness. Today, one built-in skill is all this agent needs.
Step 3 · Ground & test (≈30 min — the heart of F3)
Upload the guideline (f3-guideline-leave.md) into a new chat with your agent. Now the agent is grounded — it has one source of truth. Ask the two questions below in order. The first proves it can answer from the document; the second proves it's honest when the document is silent.
✅ Question 1 — a question the guideline does cover
I'm an Executive and I joined Synapxe 3 years ago. How many days of annual leave do I get this leave year, and how many unused days can I carry forward?
What a good answer looks like
18 days of annual leave — Executive grade, 2–5 years of service. Section 3.1
Up to 7 unused days may be carried forward, and must be used by 31 March of the next leave year or they're forfeited. Section 6.1Section 6.2
The answer names the sections it used, and the numbers match the document exactly.
Watch for
An answer with no citation — technically correct, but you can't verify it. Ask "which section says that?" and it should point you to 3.1.
A number that isn't in the table (e.g. "21 days"). That's a grounding failure — check the prompt is pasted correctly.
🚧 Question 2 — a question the guideline does not cover
What's Synapxe's policy on taking a 6-month unpaid sabbatical to do a full-time Master's degree?
What a good answer looks like
Says clearly: "This is not covered in the guideline." The document has no unpaid-leave or sabbatical policy.
Points the user to the right place: People & Culture (HR), or Head of Department for exceptions. Section 8
Does not invent a duration, an eligibility rule, or a number.
This is the key moment
A general chatbot would happily make up a plausible-sounding sabbatical policy. A grounded agent refuses — because the honest answer is more useful than a confident wrong one.
If your agent invents a policy here, that's the teaching moment: tighten the "not covered" rule in the system prompt and try again.
✅ Checkpoint: A cited answer to Q1 (criterion #3) and an honest "not covered" to Q2 (criterion #4). If time is tight, protect this step — grounding is the most important thing anyone learns today.
Optional 30-second test: ask "Can I cash out 5 unused annual leave days this year?" — a good agent answers no, only on cessation of employment and cites Section 6.3. It's a nice example of a subtle answer that is in the document.
Step 4 · Build the briefing (≈4 min)
A grounded answer is useful. A grounded answer turned into a polished, on-brand document you can send is what saves real time. Enable the Document Builder skill, then ask the agent to package what it found into a short staff briefing.
PROMPT — ask the agent to build the briefing
Using only the guideline, create a one-page staff briefing titled "Annual Leave — Quick Guide for Staff".
Structure it as:
- A one-line summary
- "Your entitlement" — the annual leave table, in plain words
- "Carrying forward days" — the carry-forward and expiry rules
- "How to apply" — the steps and the notice periods
- A closing line: "Always check the full guideline (GL-HR-014) or ask HR before acting."
Keep it under one page, in clear language for all staff. Cite the section beside each point. Format it as a clean, professional Synapxe document.
The Document Builder skill applies Synapxe's template and styling so the output looks like an official document — not raw chat text. Review it: the facts should trace back to the guideline sections you cited in Step 3.
✅ Checkpoint: A plain-language, brand-compliant briefing that a colleague could actually read. That's success criterion #5 — and all five are now done.
Human still checks. The briefing is a strong first draft, not a signed-off document. Before it goes anywhere, a person confirms the facts, the tone, and the classification. Always verify responses.
Debrief · what just happened
You built a working agent in four moves. Step back and notice what made it trustworthy:
📄
It answered from your document
Not the open internet. The same agent with a different guideline becomes a different assistant — the pattern travels.
🚧
It admitted what it didn't know
"Not covered" is a feature, not a failure. Honesty about gaps is what makes an agent safe to rely on.
📝
It produced something you can use
A grounded answer became an on-brand briefing — the difference between "interesting" and "useful".
🛡️
A human stays in the loop
The agent drafts and cites; a person verifies and decides. That's the trust posture for Restricted, Sensitive (Normal) work.
🎯 Next: make it yours. You just built the shared agent. In F4 — the Agent Design Canvas, you'll design an agent for a real task from your own team, using this exact pattern as the template.
This is the fictional document you'll upload in Step 3. Every figure and rule is invented for training — the section numbers are what the agent should cite. Download it, or copy the text into a document to upload.
SYNAPXE — STAFF LEAVE & TIME-OFF GUIDELINE
Document ID: GL-HR-014 | Version: 3.2 | Effective: 1 April 2026
Owner: People & Culture (HR) | Classification: Restricted, Sensitive (Normal)
Note: Synthetic training document. All figures, names and rules are fictional.
Leave records are personal data, handled under the PDPA.
1. PURPOSE & SCOPE
1.1 This guideline explains the paid leave and time-off available to Synapxe staff, how to apply, and who approves each type of leave.
1.2 It applies to all confirmed full-time employees. Contract, part-time and probationary staff are covered by their letters of appointment, which take precedence.
1.3 Where this guideline is silent, refer to your letter of appointment or contact People & Culture (HR). See Section 8.
2. DEFINITIONS
- Working day: a day you are normally scheduled to work, excluding weekends and gazetted public holidays.
- Service year: a completed 12 months of continuous employment from your confirmed start date.
- Leave year: 1 April to 31 March of the following year.
- HR portal: the self-service system used to apply for, approve and track all leave.
3. ANNUAL LEAVE
3.1 Entitlement (per leave year), by grade and completed years of service:
Grade band | <2 yrs | 2-5 yrs | 5-10 yrs | 10+ yrs
Associate / Officer | 14 | 16 | 18 | 21
Executive / Senior Executive | 16 | 18 | 21 | 25
Manager & above | 18 | 21 | 25 | 28
3.2 Pro-ration: new joiners and mid-year grade changes receive leave pro-rated to completed months of service, rounded up to the nearest half-day.
3.3 Minimum notice: 3 working days or fewer -> apply at least 3 working days ahead. More than 3 working days -> apply at least 10 working days ahead.
4. MEDICAL LEAVE
4.1 Outpatient sick leave: up to 14 paid days per leave year, with a medical certificate (MC).
4.2 Hospitalisation leave: up to 60 paid days per leave year (inclusive of the 14 outpatient days).
4.3 Submitting an MC: upload to the HR portal within 2 working days of the first day of absence. Sick leave without a valid MC may be recorded as annual leave or unpaid absence.
5. APPLYING FOR LEAVE & APPROVAL AUTHORITY
5.1 Apply and get approval through the HR portal. Verbal or email approvals do not count.
5.2 Approving authority by duration:
Up to 5 working days -> Reporting Manager
6 to 15 working days -> Head of Department
More than 15 working days -> Head of Department AND Divisional Director
5.3 Before approving leave over 5 working days, the manager confirms work coverage is arranged.
6. CARRY-FORWARD & ENCASHMENT
6.1 Carry-forward: a maximum of 7 unused annual leave days into the next leave year.
6.2 Expiry: carried-forward days must be used by 31 March of the following leave year. Days beyond the 7-day cap, or not used by 31 March, are forfeited.
6.3 Encashment: unused annual leave is NOT encashable during employment. The eligible balance is encashed only upon cessation of employment.
7. OTHER PAID LEAVE
7.1 Childcare leave: for eligible parents of a Singapore Citizen child under age 7, per prevailing MOM provisions.
7.2 Compassionate leave: up to 5 working days on the death of an immediate family member.
7.3 Marriage leave: 5 working days, taken within 3 months of a first legal marriage.
7.4 Examination leave: up to 5 working days per leave year for a course formally sponsored or approved by your department.
8. ESCALATION & EXCEPTIONS
8.1 Requests outside this guideline, or exception requests, go in writing to your Head of Department, who may consult People & Culture (HR).
8.2 For questions about your specific entitlement, or a situation not addressed here, contact People & Culture (HR) through the HR portal before assuming an outcome.
9. DATA HANDLING & CLASSIFICATION
9.1 This guideline is Restricted, Sensitive (Normal) — internal circulation only.
9.2 Leave records (including MCs and reasons for leave) are personal data. Access is limited to the employee, approving managers and authorised HR staff, per the PDPA.
9.3 When using an AI assistant to interpret this guideline, always verify responses against the source text, and do not upload documents above Restricted, Sensitive (Normal).
End of document — GL-HR-014 v3.2. Synthetic content for training use only.
Deliberate gaps (for the facilitator): this guideline has no policy on unpaid leave, sabbaticals, study leave, or working from home. Those gaps are what make the "not covered" test in Step 3 work. If you swap in a different document, make sure it still has at least one clear gap.
🎤 Facilitator notes
Run-of-show
F3 is a 60-minute block. The first 10–12 minutes are the instructor demo (AgentSea in Action); the rest is this shared build. Keep everyone on the same step — this is a warm-up, not a race.
Block
Facilitator focus
12–22
Walk the room through Configure. Have the system prompt on screen; read the grounding and "not covered" lines aloud.
22–26
Everyone enables Document Analysis. Quick check that no one has extra skills switched on.
26–48
Upload + Question 1. Circulate. Get everyone to a cited answer before moving on.
48–56
Question 2. This is the payoff — pause here and discuss what a general chatbot would have done.
56–60
Step 4 briefing + debrief. If short on time, demo Step 4 from the front rather than dropping it.
Protect the grounding test
If the clock slips, cut Step 4 to a live demo — but never skip Question 2. The contrast between a cited answer and an honest "not covered" is the single learning outcome that carries into every track.
Common situations & how to handle them
If you see…
Do this
Agent answers Q1 with no citation
Ask it "which section says that?" live. Reinforce that an answer you can't trace is an answer you can't trust.
Agent invents a sabbatical policy for Q2
Celebrate it as the teaching moment. Show the "not covered" rule in the prompt; have them re-ask. Grounding is a prompt behaviour, not magic.
Agent gives a wrong number (e.g. 21 not 18)
Usually a mis-paste or the file didn't attach. Re-upload the guideline and confirm the prompt is intact.
Someone asks about their real leave
Redirect gently — this is synthetic data for learning the pattern, not an HR service.
Scale & safety
One synthetic document and one prompt keeps a room of ~50 in sync. Share the guideline file ahead of time so no one waits on a download.
Keep the language roadmap-honest: connectors, knowledge spaces and shareable custom skills are concepts subject to platform readiness, not features to promise dates for.
Reinforce the governance line verbatim: AgentSea handles up to Restricted, Sensitive (Normal) information — always verify responses.
No Amazon Q framing. AgentSea is the day-one tool; Bedrock, AgentCore, Strands and Kiro are the deeper stack for later tracks.
Bridge to F4: close by asking "what document in your job would you point an agent at?" — that question is the on-ramp to the Agent Design Canvas.