Module F2 · Prompting your agent

The CRAFT of
Prompt Engineering

Context · Role · Action · Format · Trust

Synapxe · Agentic AI Foundations · Amazon Quick

A PROMPT WORTH ACTING ON

Module 1 · The running example

One quarter. One leave request.

Officer, 4 years' service I'd like 12 working days of annual leave, starting in 6 working days. My manager said yes by email. I'll have 9 unused days at year end: can I carry them all forward? A deployment stopped me taking them, so can the extra be paid out?

Notice Approver Email approval Carry-forward Pay-out Restore lost days

Source of truth: GL-HR-014 Staff Leave & Time-Off Guideline v3.2 (synthetic).

Looks routine. It is not: seven things to check.

Module 1 · Why prompts matter

The reliability gap

Lazy prompt

"Is this leave request OK?"

What comes back

Yes. An Officer with 4 years of service gets 16 days, so 12 is well within entitlement. Your manager approved it, so you are all set. Unused leave can usually be carried forward, and many organisations pay out the balance.

One thing right: 16 days. The one nobody worried about.

Notice: needs 10 days, has 6§3.3 Approver: Head of Department§5.2 Email does not count§5.1 Carry-forward cap is 7, not 9§6.1 No pay-out during employment§6.3 Restoring lost days: not covered§8.1

Not gibberish: a plausible answer that quietly skips the rules.

This is the emotional centre of the module. Read the left-hand answer aloud in a helpful, confident tone; it sounds like a good colleague. Then walk the six chips on the right as they pop in. Notice period: more than 3 working days needs 10 working days' notice (section 3.3), and this request gave 6. Approver: 12 working days is in the 6 to 15 band, so it needs the Head of Department, not the reporting manager (5.2). Email approval is not recorded; leave goes through the HR portal (5.1). Carry-forward is capped at 7 days, not 9 (6.1). Unused leave is not encashable during employment (6.3). And restoring days lost to a deployment is simply not covered; exceptions go in writing to the Head of Department (8.1). The line to land: it did not invent a fake policy. It answered the easy question and stayed silent on six hard ones, and it told a colleague they might cash out leave that Westmere does not encash. That kind of error reaches a person before it reaches you. Then: every slot we are about to add closes one of those gaps.

Module 1 · The framework

CRAFT: five slots

Miss one and the agent guesses. Press space to light each slot.

CONTEXT

What it needs to know

ROLE

Who it is being

ACTION

What to produce

FORMAT

What shape it takes

TRUST

What it must never do

Press space five times, one slot per press. 1. CONTEXT: give it the actual document, not a description of it. Skip this and it fills the gap from general knowledge. 2. ROLE: who is doing the reading? A leave administrator notices different things than a general assistant. 3. ACTION: name the deliverable. "Look at this" is not an action. 4. FORMAT: what shape should the answer take? This is what makes it reusable next month. 5. TRUST: pause here. This is the slot almost nobody writes, and it is the one that would have caught the misses on the previous slide. Trust means three moves: cite the section, say "not covered" when the document is silent, and name who decides. Closing line: two slots earn most of the quality, C and T. C stops it inventing. T makes it citable and lets it say "I don't know". Amazon Quick link: Context lives in a Space; Role, Action, Format and Trust together become the Chat Agent's persona instructions. We come back to that near the end.

C · Context

Give it the document and the facts

Four kinds of context. Press space for each.

Domain

GL-HR-014 v3.2
leave year 1 Apr to 31 Mar

Data

Officer, 4 yrs, 12 days,
6 days' notice, 9 unused

Situation

Leave year closes 31 Mar,
team mid-deployment

Constraints

Working days, not calendar days;
cite the clause

The guideline lives in a Space. Today's facts go in your message.

This is the most important slide for healthcare admin users. Without context the agent has nothing to check against. Ask the room: could YOU judge this request without the guideline in front of you? No. Neither can the agent. Reveal each kind one at a time. DOMAIN is the easiest to forget, because you know your world and the agent does not. DATA is most important for accuracy: without it the agent fills in plausible-sounding numbers. SITUATION shapes urgency: "the leave year closes in three weeks" changes the advice. CONSTRAINTS is the quiet one: "working days, not calendar days" is exactly the kind of thing that silently corrupts an answer. What each omission costs: skip Domain and you get generic answers; skip Data and you get invented facts; skip Situation and you get the wrong urgency; skip Constraints and you get 12 calendar days instead of 12 working days. Amazon Quick point: domain context belongs in a Space. Add GL-HR-014 once and every conversation has it; you do not paste it each time. A Space's description is an optional label for people, not instructions. Ask for one example from the room's own work for each kind.

R · Role

A role changes what the answer notices

Same facts, different questions. On our request, a leave administrator checks notice, approver and portal record.

R · Role

The role formula, and the "because"

Title

Procurement officer

Experience

10 years, public sector

Specialty

Facilities services

Trait

Reads for risks first

The persona line

"Act as a People & Culture leave administrator. You are precise about notice and approval authority, because leave approved by the wrong person leaves a coverage gap on a live team."

Quick rewrites persona instructions, so the idea matters more than the exact words.

Walk the formula one ingredient at a time. TITLE anchors the agent in a specific role. EXPERIENCE adjusts depth and confidence. SPECIALTY narrows the vocabulary. TRAIT shapes judgement calls and is the most under-used ingredient. Then the part that does the real work: the behaviour clause. Title plus experience plus specialty gets you vocabulary. It is the "because" clause, the behaviour and the reason for it, that changes what the agent actually looks for. Read the persona line aloud and point at the "because". "Thorough and risk-averse" and "efficiency-focused" produce different recommendations from the same data. Quick point, stated carefully: Quick rewrites persona instructions in its own words, so your exact phrasing is not preserved. Keep the role short and clear, and put any rule that must survive word for word into a reference document (we cover that shortly). Ask: which persona would you use for a board paper on a new vendor? A compliance officer. For an internal process review? An operations manager.

A · Action

Name an artefact, not an activity

"Is this leave request OK?"

Assess it against the guideline and draft the reply. Say what can proceed, what must change, what to escalate.

"Summarise this report"

Summarise it in 5 bullets: bed occupancy, staffing, patient wait times.

"Help me with this data"

List the top 5 vendor invoices by total amount.

"Review it" is an activity. "Draft the reply" is an artefact. Add the judgement you want.

F · Format

Same facts, two shapes

"What happened with our vendor SLA this quarter?"

No Format slot

Wall of text. Key number buried.

With a Format slot

Summary 97% → 94%

Cause 6.2 h response, target 4 h

Action Performance notice

Owner Procurement, 5 days

Scannable. Number first. Owner and date.

F · Format

Five things you can ask for

Sections

Summary · Findings · What must change

Table

Item | Guideline says | Clause | Status

Length

One page, under 400 words

Style

Write for the officer, not HR insiders

File type

Save as Markdown (.md)

Ask for Markdown when you will reuse it. Next month is an edit, not a retype.

T · Trust

The slot almost nobody writes

Cite the section

Say "not covered"

Name who decides

Answer ONLY from GL-HR-014. Cite the clause for every finding.
If it is not addressed, write "NOT COVERED IN GL-HR-014" and route it to §8.1.
Never fill a gap with general HR practice.

An agent that can say "not covered" is worth ten that always sound sure.

Slow down here. This is the governance slot, and for this room it is the one that makes AI usable at all. Four moves are packed into it: (1) restrict the source ("answer ONLY from GL-HR-014"), (2) force citations (the clause for every finding), (3) give it the exact words for "I don't know" AND the route for what to do next ("NOT COVERED IN GL-HR-014", then section 8.1), (4) name the specific things it must never do, for example state an entitlement figure that is not in the section 3.1 table, or advise on encashment beyond what section 6.3 says. The line to voice: we are not just telling it to be careful. We are giving it the WORDS to use and WHERE TO SEND THE QUESTION. "I don't know" is only useful if it comes with a next step. Data caveat: everything in this workshop is synthetic. Before anyone loads a real document into Quick, confirm with the DNA / IT team what is permitted. You will test this behaviour deliberately in F3: one question the guideline covers, one it does not.

T · Trust

What a good "I don't know" looks like

Can I restore the days I lost to the project deployment?

Searching GL-HR-014…

§1§2§3§4§5§6§7§8§9

NOT COVERED IN GL-HR-014

Ask your Head of Department in writing (§8.1)

A gap with a next step: that is what makes it safe to put in front of a colleague.

T · Trust

Plausible is not the same as sourced

"What is the carry-forward limit?"

No grounding rules

Staff can typically carry over around 5 days. Anything more is usually forfeited.

"Typically" and "usually" are guessing in a suit.

Grounded in the Space

Carry-forward maximum: 7 days [§6.1]

Use by 31 March [§6.2]

Pay-out: not during employment [§6.3]

Every claim points at a clause.

In Quick, switch Web search off (under + in the chat box) to stay grounded.

Read the left version aloud: it sounds completely reasonable. But it says "around 5 days" when the guideline says 7. The model made up a plausible number from its general knowledge of what leave policies typically look like. The right version cites the section and gives the exact figure. Ask: if someone planned around the left answer, who is responsible for the error? The four grounding rules, as a copy-paste baseline for any prompt where accuracy matters: (1) base your answer ONLY on the provided documents; (2) after each claim cite the document and section; (3) if it is not in the documents, say so in the exact words ("not available in the provided documents" or NOT COVERED); (4) do not use outside knowledge. Rule 3 is the most important. Quick point: your documents live in a Space, so the agent always has them. The grounding rules plus switching Web search off, under the plus button in the chat box, keep it to those documents and away from the open internet. Use this pattern for policy lookups, compliance checks, contract review and any HR query.

CRAFT · Assembled

All five slots, one prompt

CSource of truth: GL-HR-014 v3.2. Officer, 4 years, 12 working days, 6 days' notice, email approval, 9 unused days.

RAct as a People & Culture leave administrator, precise about notice and approval authority, because the wrong approver leaves a coverage gap.

AAssess the request and draft the reply the HR officer will send. State what can proceed, what must change first, what to escalate.

FOne page, under 400 words. Findings table: Item | Guideline says | Clause | Status. Save as .md

TAnswer ONLY from GL-HR-014. Cite the clause. If not addressed, write NOT COVERED IN GL-HR-014 and route to §8.1.

Written once, this becomes a Chat Agent the whole team reuses.

CRAFT · The result

What CRAFT caught

ItemGuideline saysClauseStatus

Entitlement, 12 daysOfficer, 2 to 5 yrs: 16 days§3.1Within

Notice given: 6 daysOver 3 days needs 10 working days§3.3Short

Approver: manager6 to 15 days: Head of Department§5.2Wrong level

Approved by emailHR portal only; email not recorded§5.1Not valid

Carry forward 9 daysMaximum 7; the rest is forfeited§6.1–6.22 forfeited

Pay out the extra daysNot encashable during employment§6.3Not allowed

Restore lost daysSilent. Written request to the HOD§8.1Escalate

The honest last row is not yes or no: silent, and here is who decides.

Same agent, same guideline, same 30 seconds. The difference is entirely in the prompt. Walk the table row by row, quickly; the audience has seen most of it, so the value is the completeness of the contrast with the lazy answer. Spend the time on the last row. This is the single most important habit in the workshop: an agent that can say "not covered, and here is who decides" is safe to put in front of a colleague. One that always produces an answer is not. Check any row yourself in the guideline: 3.1 is the Officer 2 to 5 years band (16 days); 3.3 notice; 5.1 portal; 5.2 approving authority; 6.1 and 6.2 carry-forward and expiry; 6.3 encashment; 8.1 exceptions. Bridge: CRAFT lays out the prompt. Now two techniques that ride on top when the stakes rise: showing its work, and looking at the same facts from several roles.

Module 2 · Chain-of-thought

Ask it to show its steps

Without

Approve this SGD 85K equipment request?

"Yes, it seems reasonable."

No maths. Not auditable.

With "think step by step"

STEP 1Budget
SGD 2.4M

STEP 2Left after spend
SGD 0.6M

STEP 3Free after committed
SGD 180K

STEP 4Request is
47% of free

RESULTAPPROVE: within threshold

If it could be audited, make it show the reasoning. The trail is your record.

Ask: if a finance officer gave you just "Yes, it seems reasonable", would you approve the budget? No. You would want to see the maths. That is what chain-of-thought does, and the example is illustrative: annual budget SGD 2.4M, SGD 1.8M spent so SGD 600K left, SGD 420K already committed so SGD 180K free, and an SGD 85K request is about 47% of that. Critical equipment, so approve within the discretionary threshold. The dot travels the chain: each step lights up in order, which is how the reasoning is laid out in the reply. Asking for the steps does more than expose them; it tends to make the answer more careful, because the maths has to be written down before the conclusion. For auditable decisions in healthcare the reasoning trail IS your documentation. If you are asked to justify an approval, you can show the chain that led to it. Only the instruction changes: add "think step by step".

Module 2 · Chain-of-thought

Four ways to ask for reasoning

Zero-shot

"Think step by step"

Quick sums, budget checks

Few-shot

"Here is one worked example"

Same steps every time, such as leave assessments

Step-back

"First list the key requirements"

Policy and compliance checks

Self-consistency

"Solve three ways, report the majority"

High-stakes choices

Zero-shot for every day. The others when consistency or the stakes rise.

Four tools in the toolkit, each for a different situation. Zero-shot is the everyday tool: literally add "think step by step". Few-shot is for when you need consistency across many similar assessments; you show one worked example and the agent follows the pattern. Step-back is for complex policy analysis: first ask the agent to state the key requirements (for example, the main requirements of an infection-control circular), then assess your own practices against them, so it never says "some areas need improvement" without saying against what. Self-consistency is for when being wrong is costly: evaluate several independent ways and take the majority. Healthcare admin rule: any decision that could be audited should use one of these. The next two slides show few-shot and self-consistency in action.

Module 2 · Few-shot

One worked example sets the pattern

Your example: Staff A

Balance: 14 − 10 = 4 days

5 days asked > 4 left ×

Backup: none named ×

DECLINE · offer 4 days

Agent output: Staff B

Balance: 14 − 6 = 8 days

3 days asked < 8 left ✓

Backup: confirmed ✓

APPROVE

Show the pattern once. It repeats it for the next hundred.

This is the most powerful reasoning technique for repetitive business processes. The example does not just show the answer; it shows the exact reasoning pattern you want. The agent follows that pattern precisely for Staff B (the numbers are illustrative and are not taken from GL-HR-014). Ask: how long would it take to train a new HR admin to follow this process? Days. With a worked example the agent picks it up from one. This is perfect for Amazon Quick: put the worked example into the agent's instructions and every leave assessment follows the same logic. Compare with the zero-shot version: "Should we approve this leave application?" gets a vague yes with no balance check and no coverage check; "appears sufficient" is not a metric. If HR asked you to justify an approval, which version would you want in your records?

Module 2 · Self-consistency

Three angles, one vote

If the angles disagree, flag it for committee review.

Module 2 · Several roles

Three meetings in one prompt

Policy change: extend visiting hours from 8pm to 10pm.

When the readings disagree, that is your signal to ask a person.

In real organisations you would need three meetings with three departments. With several roles in one prompt you get all three readings at once, each with its own concerns, a recommendation and one metric to monitor, followed by a synthesis. Finance says expensive, clinical operations says beneficial with caveats, compliance says allowed with safeguards, and the synthesis combines them into a practical pilot. Figures are deliberately left out; ask the agent to use only the numbers you supply. Same trick as a role: the persona decides what gets emphasised. Compliance, operations, finance and a clinical engineer can each read the same vendor SLA data and recommend very different next steps. All are valid; the choice depends on the audience. Ask: which would you show the governance committee, which the operations team, which the CFO? Remember the earlier rule: if the Finance reading says fine and the Procurement reading says gaps, the disagreement is the signal to ask a person before signing.

Module 2 · Audience

Same data, four audiences

10-minute practice: write prompts for two of these. Compare the tone.

Module 3 · Amazon Quick

Where each slot lives in Quick

In F3 you turn a brief of exactly this shape into a Chat Agent.

This is the bridge between prompt engineering and agent building. Everything you learned is packaged permanently. Context goes into a Space: add the guideline once and every conversation can search it. Role, Action, Format and Trust together become the Chat Agent's persona instructions. Anything that must survive word for word (a citation rule, an exact refusal line) goes into a reference document. Facts to say plainly. A Space's description is an optional label for people; it is not instructions and does not steer the agent. Persona instructions are rewritten by Quick with generative AI, so exact wording is not preserved. Reference documents are preserved verbatim. "Update preview" only refreshes the test pane; "Launch chat agent" saves and publishes. After editing an agent, start a NEW conversation, because a chat remembers its history. Platform quirk: a Space created a moment ago can be missing from the picker until the browser is reloaded. Reload before assuming it failed.

Module 3 · Amazon Quick

One template for every agent

## RoleYou are [title] for [organisation], who [trait] because [reason].R

## TaskWhen given [input], you will 1) [step] 2) [step] 3) produce [artefact].A

## OutputSections, table columns, length, file type.F

## RulesOnly from the Space. Cite the section. Else: NOT COVERED. No clinical advice.T

## ExampleOne worked example of a good answer (optional).+

Every technique from today has a home here. Context stays in the Space.

This is the master template that connects everything from today. Role is the persona with its "because" clause. Task is the Action slot: a numbered list ending in an artefact. Output is the Format slot. Rules is the Trust slot, plus the healthcare guardrails on the next slides. Example is where a worked example (few-shot reasoning) goes. Context is deliberately not in the text: it lives in the Space. Worked example for the room, a vendor assessment agent: Role: Procurement Assessment Specialist with 10 years in public healthcare procurement. Task: extract commercial terms (price, SLA, duration, penalties), compare against standard requirements, produce a GREEN, AMBER or RED assessment. Rules: only the uploaded document, amounts in SGD, no clinical advice, no personal details, "NOT IN DOCUMENT" when something is missing, human review for large contracts. Output: Executive Summary, Commercial Terms table, Risk Assessment, Recommendation, Next Steps. Those instructions plus a Space holding your proposals make a reusable assessment agent that any team member can run. Remember the Quick caveat: rules that must be quoted exactly belong in a reference document, because persona instructions are reworded.

Module 3 · Amazon Quick

Some rules must not be reworded

Persona instructions

If not covered, write: NOT COVERED IN GL-HR-014

Quick rewrites it

Tell the user the guideline does not seem to cover this.

Reference document

If not covered, write: NOT COVERED IN GL-HR-014

Kept word for word

If not covered, write: NOT COVERED IN GL-HR-014

Exact wording that matters goes in a reference document.

An important Quick fact, shown with one rule travelling two routes. Persona instructions are rewritten by Quick with generative AI, so the wording you typed is not preserved and a precise phrase may come back softened. Reference documents keep their wording verbatim. AWS's own guidance for agent context says critical guidelines that must remain unchanged belong in reference documents. So: the idea of the role goes in the persona instructions; the exact refusal line, the citation format and any "never say X" rule go in a reference document attached to the agent. Do not claim anything stronger than that: the top lane shows what CAN happen, not what always will. Link back to Trust: this is how "NOT COVERED IN GL-HR-014" stays exactly that. Also tell the room again that the Space description is a label, not instructions.

Module 3 · Amazon Quick

Six guardrails in every agent

NEVER

Give clinical advice

Put personal details in outputs

Make the final approval

ALWAYS

Cite the source

Flag what is not in the document

Send big decisions to a person

Say what it must not do. Defaults are helpful and over-eager.

These guardrails protect three things: patients (no clinical advice), the organisation (no autonomous approvals), and good practice with personal data (keep identifiers out of outputs; for what your organisation requires, ask your DNA / IT team and follow your data-protection rules). Reveal the left side first, the three NEVERs: clinical advice (the agent is not a clinician), personal details such as names and ID numbers, and final approvals (AI assists, humans decide). Then the three ALWAYS: cite sources so every claim is checkable; flag uncertainty with a fixed phrase such as NOT IN DOCUMENT; and send decisions above your approval limit to a named person. Negative constraints feel counterintuitive but work. Most people only say what TO do. The agent has strong defaults: it wants to be helpful and comprehensive, so it may add medical advice to a procurement answer. Useful examples: "Do NOT provide clinical recommendations", "Do NOT reference data outside the provided documents", "Do NOT use phrases like 'it depends'", "Do NOT exceed 300 words", "If a detail is not in the document, write NOT IN DOCUMENT". Put these in every agent as a baseline; remember that exact wording belongs in a reference document. The Agent Design Canvas in F4 pre-fills them.

Module 3 · Amazon Quick

Which surface for which job?

Chat on a Space

A one-off question over your documents

Chat Agent

The same task, repeatedly, shared with the team

Quick Flow

Runs without you, on a schedule or a trigger

Quick Research

Plans its own multi-source investigation

Most of the value is a Space plus a Chat Agent. Automate later.

This is the practical "where do I start" slide. Emphasise the rule of thumb: people over-reach for automation before they have a working agent. Start with a Space and a chat. Formalise it into a Chat Agent once the shape is clear. Then automate. Examples: a plain chat on a Space answers "what does the procurement policy say about single-source awards?". A Chat Agent suits vendor assessments, leave queries and policy briefings. A Quick Flow suits a Monday status brief or triaging an arriving document, and a Flow can invoke a Chat Agent you already built, so the work compounds. Quick Research suits a market scan, a supplier risk review or a regulatory landscape. Quick Flows are available on all plans, but whether you can create one depends on your role permissions. Quick Automate exists above all this for enterprise processes with approval gates, but it is Enterprise tier only, so it is an instructor demo here and a conversation with DNA / IT.

Module 4 · Better results

Draft, score, revise

Three extra lines in your prompt. The agent checks its own work first.

One of the most useful optimisation techniques for business users because it needs no technical skill, just three extra lines in the prompt. Step 1 DRAFT: write the vendor performance summary from the data provided. Step 2 SCORE: rate your draft on completeness (all required sections?), grounding (every claim cites data?) and actionability (specific next steps?), each 0 to 5. Step 3 REVISE: if the total is below 12, rewrite to fix the lowest-scoring area, at most 2 revisions, and output only the final version. The scoring rubric is what makes it work, because it forces the agent to check its work against specific standards you chose. "Max 2 revisions" prevents endless loops. We are not claiming a measured improvement figure; try it on your own task and see whether the first draft you read is better. Pair it with Trust: a score for grounding is a natural place to ask whether every claim cites a clause.

Module 4 · Better results

Break big tasks into small steps

One giant prompt

Extract termsCompareFind risksWrite memoBuild tableRank vendors

Six jobs, shallow on each.

Four chained prompts

Check and correct after each step.

In Quick, that is four messages in one conversation instead of one.

Module 4 · Wrap-up

Seven mistakes everyone makes

Press space (or click a card) to flip it to the fix.

Kitchen sink

Five tasks crammed into one prompt.

Fix

One task per prompt. Chain the results.

Blank canvas

No example, so it guesses your format.

Fix

Show one or two examples of the output.

Trust fall

No grounding, so confident guesses.

Fix

"Answer ONLY from the provided documents."

Vague ask

"Analyse this": what, how, for whom?

Fix

Name the audience, format and length.

One-shot wonder

Expecting perfect on the first try.

Fix

Plan two or three refinement turns.

Copy-paste trap

A generic prompt, never adapted.

Fix

Add your domain, data and constraints.

#7 · Healthcare

No guardrails

No clinical-advice or personal-data rules.

Fix

Always add the "do not" rules.

Your defence

A saved Chat Agent holds all seven fixes.

This slide always gets a reaction because everyone recognises at least three or four of these. Reveal them one at a time (each space press flips a card) and after each one ask: who has done this? Show of hands. #1 The Kitchen Sink: the most common. People ask the AI to analyse, summarise, recommend AND format at once; each task gets diluted attention. #2 Blank Canvas: no example means the agent invents a format you did not want. #3 Trust Fall: the most dangerous in healthcare, because the agent sounds confident even when it is guessing; always require "ONLY from the documents". #4 Vague Ask: "analyse this" is not a prompt, it is a wish. #5 One-Shot Wonder: the first output is a draft, not a deliverable. #6 Copy-Paste Trap: generic prompts work generically, so add YOUR context. #7 No Guardrails (healthcare): forgetting "no clinical advice" or "no personal details" can have real consequences. The defence reveals last: the fix is permanence. Bake the seven rules into a Chat Agent's instructions (with exact wording in a reference document), launch it in Quick, and you stop relying on memory.

Hands-on · Prompt exercises

Choose your exercise

Vendor proposal assessment

Best for procurement, operations, compliance.

Output: a reusable GREEN / AMBER / RED template

Monthly operations report

Best for finance, administration, HR.

Output: a reusable executive summary template

Open the prompt exercises

Not sure? Start with the vendor exercise. Allow 45 minutes.

Wrap-up · What you can now do

Seven things you can do now

Write a prompt with all five CRAFT slots

Ask for reasoning you can audit

Use roles for different readers

Get structured, sourced answers

Add guardrails to agent instructions

Pick the right Quick surface

Turn a prompt into a reusable template

Twenty minutes once buys a template your team reuses.

Walk through each outcome. The last one is the bridge to the build. Ask: who feels confident they could write the instructions for their most common task? Most hands should go up. Likely worries, and the answer to each. "I don't want to write long prompts every time": build the template once and launch it as a Chat Agent. "I'm not sure my prompt is good enough": use the draft, score, revise loop. "My team needs consistent outputs": a shared Chat Agent plus a structured format gives the same shape every time. "How do I know it isn't making things up?": the Trust slot, grounding rules and "not covered". "What about personal data and compliance?": guardrails in every agent's instructions, and ask your DNA / IT team what may be loaded. The prompt templates built today are real deliverables you can turn into a Chat Agent in Quick immediately.

Next · Module F3

Now build it in F3

Your CRAFT brief becomes a Chat Agent in Amazon Quick.

Create a Space

Add GL-HR-014 with + Add, then Upload files.

Make a Chat Agent

Start from the + Blank card. Use your brief.

Test it twice

One question it covers, one it does not.

Open the F3 lab Back to the workshop hub

Think of a weekly task that takes too long and follows a pattern.

Close with: today you learned to communicate with AI effectively. Next, in F3, you build it. Create a Space and add the guideline, create a Chat Agent from the + Blank card using your CRAFT brief, then test it with two questions: one the guideline covers and one it does not. If the agent invents an answer to the second, strengthen the Trust slot and put the exact wording in a reference document. Remind the room of two platform facts: a Space created a moment ago may not appear in the picker until the browser is reloaded; and after you edit an agent, start a new conversation to test it. The F3 lab and the workshop hub are one click away from the buttons on this slide. Then plant the seed for F4, the Agent Design Canvas: what task does your team do every week that takes too long and follows a predictable pattern? That is your agent candidate. Thank everyone.

The CRAFT ofPrompt Engineering

One quarter. One leave request.

The reliability gap

CRAFT: five slots

Give it the document and the facts

A role changes what the answer notices

The role formula, and the "because"

Name an artefact, not an activity

Same facts, two shapes

Five things you can ask for

Sections

Table

Length

Style

File type

The slot almost nobody writes

Cite the section

Say "not covered"

Name who decides

What a good "I don't know" looks like

Plausible is not the same as sourced

All five slots, one prompt

What CRAFT caught

Ask it to show its steps

Four ways to ask for reasoning

One worked example sets the pattern

Three angles, one vote

Three meetings in one prompt

Same data, four audiences

Where each slot lives in Quick

One template for every agent

Some rules must not be reworded

Six guardrails in every agent

Which surface for which job?

Draft, score, revise

Break big tasks into small steps

Seven mistakes everyone makes

Kitchen sink

Blank canvas

Trust fall

Vague ask

One-shot wonder

Copy-paste trap

No guardrails

Your defence

Choose your exercise

Vendor proposal assessment

Monthly operations report

Seven things you can do now

Now build it in F3

The CRAFT of
Prompt Engineering