You take away a completed canvas and tested instructions. They are the blueprint for building this agent properly.
If none of the ten workflows is your problem
Take the last tile, Bring your own, and describe a real task from your team. People, HR and change work belongs there; there is no pre-built card for it.
Step 1 · 5 min
Pick your workflow
Select a tile to preview the agent you would be creating. Nothing is chosen until you press Design this agent.
Best candidates: a task you do weekly, that takes 30 minutes or more, and starts from a document you may upload. Short on time? Start with a Quick starter.
Quick starterssimplest · any team · start here if you’re short on time
Meeting Notes → ActionsAny team
Notes in, minutes and a traceable action table out.
45 min → ~10 minDaily
Inbox / Request TriageAny team · Service Desk
Classify by urgency with a stated reason, suggest owners, draft the routine replies.
40 min → ~10 minDaily
Leadership & Governancecommittee packs · policy and circular compliance
Committee / Board Paper Briefing PackLeadership & Secretariat
Compress a long paper set into a one-page decision pre-read per paper.
Score bidder responses against the published criteria, flag gaps, recommend.
3 hrs → ~25 minWeekly
Or bring your ownnothing above fits? this is the better option
My Own WorkflowYour function · your problem
Design an agent for a task from your own work. Answer four quick questions and the canvas fills in around your answers.
Guided setup
Define your workflow
Nothing is pre-decided here. The canvas fills in around your answers, so a vague answer gives a vague agent.
What shape is the work? pick the closest — this maps to a workflow pattern
Your agent
Data classification — before you upload anything
Sample packs provided in this session are synthetic and PDPA-safe. If you swap in a real document it must contain no patient-identifiable information — and confirm the approved classification for Amazon Quick with the DNA / IT team before you upload anything real. When in doubt, don't upload — describe the document in your canvas instead and use the synthetic sample for the build. And whatever the agent produces, always verify responses against the source before acting on them.
Step 2 · 20 min CRAFT all five
Fill the canvas
Every agent is the same five parts, and the five CRAFT slots map onto the boxes one for one.
The canvas decides what goes in each box.
At Step 5 this diagram is redrawn from your own answers. A design is done when you can say what the agent does alone and what always waits for a person.
The canvas is your plan, not the agent's instructions (those come in Step 3). The pre-fill is a strawman: sharpen it, especially the guardrails.
ALet Quick interview yourecommended
Paste one prompt into Amazon Quick and it asks you five questions — one per CRAFT slot — each with lettered options you can just pick from. Then type BUILD IT and it writes the finished canvas, a pitch, and an architecture sketch. You design an agent by talking to one.
It is a long prompt — paste it as one message and send it before answering anything.
or
BFill it in yourselvesno Quick needed
Type straight into the canvas below with your team. Works offline and works if your lab has expired. Slower, and easier to leave the pre-fill unchallenged — which is the trap this step is about.
Bring the interview result backpaste what Quick produced after BUILD IT
Paste the whole reply — canvas, pitch and architecture together. We'll read the canvas into the worksheet below and keep the other two with your download.
First five headings, in order: C Data & Context · R Agent Identity · A Workflow · F The Outcome · T Guardrails. The rest is the business case, not prompt content.
AGENT DESIGN CANVAS · editable worksheet
# Agent Design Canvas — Amazon Quick Edition
> Pick a workflow in Step 1 and this canvas fills in around it.
> Or fill it in from scratch — the headings are the same either way.
> The five CRAFT slots come first, in order. C · R · A · F · T.
## Data & Context · C
- **The document analysed:** [the specific document for this task — lives in the Space]
- **Standing knowledge in the Space:** [the policy, template, criteria the agent always needs]
- **Classification:** [no patient-identifiable information — confirm the approved classification with DNA / IT]
## Agent Identity · R
- **Name:** [a descriptive name a colleague would understand, e.g. "ITQ Evaluation Assistant"]
- **Role:** [persona formula: Title + Experience + Specialty + Behaviour]
e.g. "Senior procurement analyst with 10 years of healthcare tender evaluation experience, methodical and evidence-first"
- **Purpose:** [one sentence — what does this agent help you do?]
## Workflow · A
- **Trigger:** [what starts it? e.g. "when a bidder response lands in the tender folder"]
- **Steps:**
1. [what do you give it?]
2. [what should it do with it?]
3. [what should it hand back?]
## The Outcome · F — be specific
- **Artifact:** [the actual thing produced — a 2-page memo? a table? a draft email?]
- **Structure:** [sections, columns, length — "a 5-column table plus a one-paragraph recommendation"]
- **Good enough looks like:** [how will you judge it? e.g. "every score cites a clause"]
## Guardrails · T
- **Must always:** [e.g. "quote the clause that supports each finding"]
- **Must NEVER:** [at least 3 — these matter more than the instructions]
- **Escalate when:** [at least 2 — when should it stop and hand to a human?]
- **Tone:** [who reads the output, and how formal does it need to be?]
## Amazon Quick Capabilities
- [ ] Chat agent + Space — the document analysed, held in the Space
- [ ] Space + knowledge base — the documents the agent should always know
- [ ] Quick Research — cited in-depth research report
- [ ] Quick Sight — dashboards and data analysis
- [ ] Quick Flows — package it as a repeatable automation
- [ ] Action connectors — read from or update another system
## Business Impact
- **Today:** [how long does it take, done manually?]
- **With the agent:** [expected review time]
- **Frequency:** [how often?]
- **Time saved per month:** [time saved x frequency — do the arithmetic]
## First Step After Today
- **Who else would use this?** [one name]
- **What would you need to build it properly?** [data access, a template, a decision]
What each Amazon Quick capability means
What the capabilities mean on Amazon Quick
Amazon Quick puts a chat agent in front of a Space. The agent holds the instructions and persona; the Space holds the documents, datasets and connectors it draws on. For F4, the first two are all you need — the rest are worth knowing so you can see where the design could go.
Chat agent + SpaceA custom agent with your instructions. The document, plus its standing knowledge, lives in a Space the agent searches at answer time — nothing to attach in chat.Use it today
Space + knowledge baseThe knowledge the agent should always have — the policy, the template, the criteria. Assign the Space to the agent and share it so your team works from the same context.Use it today
Quick ResearchProduces an in-depth report with citations, drawing on the Space and the web. Useful when the task needs evidence gathered, not just a document read.Design for it
Quick SightDashboards and analytics over connected datasets — for when the input is data rather than documents.Design for it
Quick FlowsPackages the task as a repeatable automation you can invoke from chat or share with the team — built in natural language or a visual editor, no coding.Design for it
Action connectorsLets the agent read from or update another system. Powerful, and the point at which governance questions get serious.Design for it
What separates a strong canvas from a weak one
What separates a strong canvas from a weak one
The outcome is concrete. "A summary" is weak. "A 5-column table plus a PROCEED/CLARIFY/REJECT line" is a specification.
The trigger is a real moment. "When a bidder response lands" beats "when needed".
The NEVER list is specific to your risk. Generic safety rules are table stakes; the valuable ones name the mistake your workflow actually makes.
Escalation is defined. An agent that knows when to stop is worth more than one that always answers.
The impact has arithmetic. "3 hrs → 25 min, 4x/week" is a business case. "Saves time" is not.
Step 3 · 12 min CRAFT assemble
Write the instructions
The canvas is for people. The instructions are what the agent runs. Each canvas slot becomes a block.
Canvas slot to instruction block
C → Your Sources
Data & Context. Names the input and the Space.
R → You are…
Agent Identity. One persona sentence.
A → What You Do
Workflow. Numbered steps.
F → Output Format
The Outcome. Artifact, length, tone.
T → four blocks: Grounding Rules, You Must NEVER, When to Escalate, How to Close
Guardrails grew the most, because everything that makes an agent safe to rely on lives here.
ALet Quick write it from your canvasrecommended
Paste one short prompt into Amazon Quick — ideally in the same chat you used for the interview, so it works from the canvas it already produced. (The prompt also carries the canvas verbatim, so a fresh chat works too.) Quick keeps every specific you named and returns the paste-ready instructions in one fenced block. Bring the reply back here.
The prompt embeds your current Step 2 canvas. Sharpen the canvas first if you want changes reflected.
or
BAssemble from the templateyou write it
Read the canvas and fill every [bracket] in the scaffold below by hand. Slower — but it drives the CRAFT-slot to instruction-heading mapping home. Also the path to take if the lab has expired.
Bring the instruction result backpaste what Quick returned to Path A
Paste the whole reply. We'll pick the paste-ready block out of it (the one that starts with You are…) and drop it into the template box below — the same box Path B populates by hand.
AGENT INSTRUCTIONS · paste this into Amazon Quick
You are [AGENT NAME] — [ROLE, using the persona formula]. # R
Your purpose is to [ONE SENTENCE from the canvas].
## Your Sources · C
- The document to analyse, in the "[SPACE NAME]" Space: [INPUT DOCUMENT]
- Standing knowledge always available to you in the same Space: [POLICY / TEMPLATE / CRITERIA]
- Use nothing else. Not general knowledge, not assumptions, not the internet.
## What You Do · A
1. [STEP 1 — what to read and extract]
2. [STEP 2 — what to compare, check or reason about]
3. [STEP 3 — what to produce]
## Output Format · F
Produce [THE ARTIFACT — be exact]:
- [Section or column 1]
- [Section or column 2]
- [Section or column 3]
Keep it to [LENGTH]. Use [TONE] language suitable for [WHO READS IT].
## Grounding Rules · T — these are not optional
- Answer only from the uploaded document and the documents in your Space.
- For every finding, quote or cite the section it came from.
- If the documents do not cover something, say "NOT COVERED IN THE SOURCE" — never fill the gap.
- If a figure is missing, say "NOT PROVIDED" — never estimate, infer or round.
- All monetary values in SGD, stated exactly as written in the source.
## You Must NEVER · T
- [From canvas — NEVER rule 1, specific to your workflow]
- [From canvas — NEVER rule 2]
- [From canvas — NEVER rule 3]
- Include patient-identifiable information in any output.
- Give medical, legal, or financial advice.
- Present your own inference as something the source document says.
## When to Escalate · T
Stop and say " This needs human review — please check with [ROLE]" when:
- [From canvas — escalation trigger 1]
- [From canvas — escalation trigger 2]
- The source documents contradict each other.
- The request falls outside the scope defined above.
## How to Close Every Response · T
End with: "Verify this against the source document before acting on it."
The mapping in detail
Slot
From your canvas
Becomes this in the instructions
C Context
Data & Context
## Your Sources — and the Space you assign in Step 4. Name the input and the standing documents; the agent cannot ground itself in something you never mentioned.
R Role
Agent Identity
The opening You are… line. Title + experience + specialty + behaviour, in one sentence.
A Action
Workflow
## What You Do — your numbered steps, read / compare / produce.
F Format
The Outcome
## Output Format — the artifact, its sections or columns, its length and tone.
T Trust
Guardrails
Four headings, because Trust is the slot that needs the most words: ## Grounding Rules, ## You Must NEVER, ## When to Escalate, ## How to Close Every Response.
Platform guardrails vs your guardrails
Platform guardrails vs your guardrails. Amazon Quick applies its own baseline protections — toxicity filters, permission-aware answers so people only see what they already have access to, and any blocked terms your administrator has configured. Those are the floor, not the ceiling. Everything that makes your agent trustworthy — grounding, citation, refusal to guess, escalation — is what you write here.
Checkpoint before you build. Your instructions should:
Name what the agent reads — the document analysed and the standing knowledge in the Space
Name a specific role and purpose, not a generic assistant
Specify the output artifact precisely enough that two people would produce the same shape
Contain at least 3 NEVER rules that are specific to your workflow
Contain at least 2 escalation triggers
Tell the agent what to do when the source doesn't cover the question
Step 4 · 8 min
Build & test it on Amazon Quick
The goal is not a finished agent. It is to see your instructions behave and to catch the one thing you got wrong.
Six moves, and what you should see
Create a chat agentUse + Blank on the Chat agents page. Name and describe it for a colleague.See: an empty agent form.
Paste your instructionsThey go in as persona and behaviour. Quick may reword them.See: the preview pane reflects them.
Check the Space came acrossUnder Knowledge sources, confirm your Space is listed. Add it only if missing.See: your Space by name.
Add 2 or 3 suggested promptsThe everyday ask, not the clever one. Three are ready below.See: prompt chips on the agent.
Test with two questionsOne the documents answer, one they do not. Nothing to attach.See: a cited answer, then a plain “not covered”.
Share itLet your partner try it cold. Remember Launch chat agent saves it.See: your partner gets the same behaviour.
Which rules survive word-for-word (use Reference documents)
Which of your rules survive word for word. Amazon Quick treats what you paste
into persona instructions as a starting direction: it expands them with
generative AI to fit the chat features and the resources you attached, and
your original wording may be changed in the process. That is fine for
R and A — a reworded persona still behaves like the expert
you described. It is a problem for the parts of T you need verbatim: the exact
refusal script NOT COVERED IN THE SOURCE, the
NOT PROVIDED rule, your NEVER list, your escalation triggers.
AWS is explicit about the fix: guidelines that must remain unchanged belong in
Reference documents, which Quick preserves exactly rather than paraphrasing.
You met that distinction in F3. So if a rule matters
enough that you would notice it being softened, put your ## Grounding Rules,
## You Must NEVER and ## When to Escalate sections into a short
.md or .txt file and upload it under
Reference documents as well. Limit 10 files, 50 MB.
Not today — eight minutes is not enough. But it is the first thing to do when you
build this properly, and it is the difference between an agent that usually refuses and one that
always does.
Pick a workflow in Step 1 — or describe your own — and three suggested prompts appear here.
Pick a workflow in Step 1 — or describe your own — and your two test questions appear here.
The five things to check
Does the output match the structure you specified — or did it drift into prose?
Does it cite the section for each finding, without being asked again?
Ask something the documents don't cover. Does it say so, or does it improvise?
Remove a figure from the input. Does it say NOT PROVIDED, or estimate?
Trigger an escalation condition on purpose. Does it stop?
If it improvises, that's the finding. An agent that invents a plausible answer is the failure mode that matters in healthcare technology — it's confident and it's wrong, and it's hard to spot downstream. When you catch it, fix the instruction that let it happen and run the test again. That loop is the actual skill this workshop is teaching.
Step 5 · 5 min
Self-check & share back
Paste your instructions and run the check. It looks for the parts that separate a deployable agent from a demo.
● Your workflow choice, team name, canvas edits and instructions are saved in this browser as you type, so a refresh won't lose them. Download the canvas before you leave — the saved copy doesn't travel to another machine.
Not a presentation. Answer three questions for the room:
What does your agent produce? Name the artifact.
What did it get wrong first? This is the most useful thing you can tell everyone else.
What would you need to run it for real? Data access, a decision, a template — this is what goes to DNA / IT.
Finished early? Step 6 turns your canvas into something you can show.
Step 6 · optional · +10 min if you finish early
Show your agent
Turn the canvas you already have into a dashboard you can show the room. Nothing here is scored, and nothing here changes your canvas.
Let Quick build it as an appabout 10 min
Copy one prompt, built from your canvas, and Amazon Quick Apps builds a dashboard of your agent that you can publish to your team. Then see what it made.
Left nav → Apps → create an app.
Paste the prompt as one message and send it. Building takes a few minutes.
Explore the preview. Then press every button you can find.
Publish it, shared with your team only.
The app shows your agent. It is not your agent.
An app in Quick cannot embed the Chat Agent you built in Step 4, so this one is a picture of the design: it connects no Space and runs no AI. Your agent's instructions from Step 3 stay the deliverable of F4.
What good looks like
Your rules are word for word. Check one must-never rule against your canvas. If it was reworded, ask Apps to use the text from the DATA block exactly.
The human checkpoint fires. Somewhere on the page, a person gets pulled in when one of your escalation triggers is met — and lets go afterwards.
Nothing is invented. Anything your canvas left open should say Not decided yet, not a made-up figure.
Apps not working, or your lab has expired?
This page can build the same dashboard itself, from your canvas, in one click. It opens in any browser and works offline.
Your canvas has edits. Your edits will be kept; this workflow’s pre-fill is held until you choose.
This is a starting hypothesis, not the answer. Step 2 will challenge it, and these are what the review looks for: a real trigger, named sources, a human gate that is a rule rather than a hope, and impact you can defend.