DAY 1 — FOUNDATIONS, INSTRUCTION DESIGN & KNOWLEDGE GROUNDING

UNDERSTANDING AGENTIC AI

Topics Covered

• From chatbot to copilot to agent: what „agentic“ precisely means, and what it does not
• The agent loop — objective, plan, act, observe, self-correct, deliver
• Anatomy of a Claude agent: model, system instruction, knowledge, tools, guardrails, human checkpoints
• Matching model to task: Claude Haiku, Sonnet and Opus, and the cost–capability trade-off
• Where agents create measurable ROI — and the categories of work they should never own

Learning Activities

Instructor presentation; live demonstration contrasting a single-shot prompt against a multi-step agentic task on the same dataset; facilitated group discussion mapping candidate processes within participants‘ own departments.

Hands-On Exercise

Task Triage Audit — Working from a list of ten recurring departmental tasks, participants classify each as prompt-suitable, agent- suitable, or human-only, and defend each classification with a two-line rationale on risk, repeatability and verifiability.

Expected Outcome

Participants can define agentic AI accurately, explain the agent loop to a colleague, and correctly identify which of their own workplace tasks are genuine agent candidates rather than AI theatre.

INSTRUCTING CLAUDE — THE CRAFT FRAMEWORK & SYSTEM PROMPTS

Topics Covered

• Format, Tone — the backbone of every reliable agent
• Writing system instructions that fix an agent‘s persona, permitted scope, and refusal conditions
• Structured output discipline: XML tags, JSON schemas, and why format consistency is the prerequisite for automation
• Task decomposition, step-by-step reasoning, and self-checking instructions for multi-step reliability
• Few-shot examples, negative constraints, and instructing an agent to say „I do not know“

Learning Activities

Demonstration of prompt-to-system-instruction conversion; prompt clinic with live rewriting; paired peer review using a supplied quality checklist; before-and-after output comparison.

Hands-On Exercise

Monthly Report Analyst — Participants convert a vague one-line request („summarise this report“) into a complete CRAFT system instruction, then test it against a real 12-page management report. Outputs are run twice to test consistency, and the instruction is refined until both runs are structurally identical.

Expected Outcome

Participants can author production-grade system instructions that produce consistent, correctly formatted output across repeated runs — the single largest determinant of agent reliability.

GROUNDING AGENTS IN ORGANISATIONAL KNOWLEDGE

Topics Covered

• Claude Projects: shared context, custom instructions, and the project knowledge base
• Preparing and structuring source material — SOPs, policies, price lists, contracts and datasets
• Context window management: what to upload, what to summarise, and what to leave out
• Reducing hallucination: source-bound answering, citation prompting, and explicit „not in source“ rules
• Data governance in practice: classification, confidentiality, and what must never enter a shared project

Learning Activities

Live demonstration of project construction; hands-on lab; adversarial testing exercise in pairs; group discussion on internal data classification policy.

Hands-On Exercise

Procurement Policy Assistant — Teams build a Claude Project loaded with a three document policy pack, then interrogate it with eight questions, two of which are deliberately unanswerable from the sources. Success is measured not by the six correct answers, but by whether the agent correctly refuses the two it cannot support.

Expected Outcome

Participants can build a knowledge grounded assistant that answers strictly from approved organisational sources, cites where its answers come from, and declines to speculate.

FROM PROMPT TO AGENT — DESIGNING THE AGENT WORKFLOW

Topics Covered

• Decomposing a business process into discrete, individually verifiable agent steps
• The Agent Brief: objective, inputs, permitted actions, success criteria, stop conditions, escalation path
• Generating real deliverables — formatted reports, spreadsheets, charts, slide outlines and draft correspondence
• Placing human-in-the-loop approval gates: which steps require sign-off and which do not
• Common failure modes: scope creep, silent errors, over-confidence, runaway loops and how to design against them

Learning Activities

Demonstration; individual hands-on build; case-study review of a failed automation; group critique of each other‘s agent briefs

Hands-On Exercise

Weekly Sales Summary Agent — Participants build an agent that ingests a raw, unclean CSV export, corrects formatting inconsistencies, calculates actual-versus-target variance by region, produces a supporting chart, and drafts a 200-word narrative summary written for a management audience. Participants then deliberately corrupt the input file to observe and document how the agent fails.

Expected Outcome

Participants can take a real recurring process, express it as a formal agent brief, build a working end-to-end agent, and articulate exactly where and why a human must remain in the loop.

DAY 2 — TOOLS, MULTI-STEP WORKFLOWS, GOVERNANCE & CAPSTONE

EXTENDING CLAUDE WITH TOOLS, CONNECTORS & MCP

Topics Covered

• Why agents need tools: closing the gap between reasoning about work and actually doing it
• Tool use and function calling — how Claude decides which tool to invoke, when, and with what arguments
• Connectors and the Model Context Protocol (MCP): linking Claude to Google Drive, SharePoint, databases and internal systems
• Working surfaces: Claude in the browser, Claude in Excel, and Claude Code for analyst workflows
• Least-privilege design: read-only versus write access, credential handling, and the blast radius of a mistaken action

Learning Activities

Live demonstration of connector setup; hands-on lab; instructor-led verification walkthrough; group discussion on IT approval and access-control requirements.

Hands-On Exercise

Connected KPI Extractor — Participants connect a shared cloud folder, instruct Claude to locate the most recent month‘s data file without being told its name, extract three specified KPIs, and write them into a formatted summary. Every figure is then manually verified against the source file, and any discrepancy is traced back to its cause.

Expected Outcome

Participants can extend Claude beyond conversation into their actual file systems and data sources, and can specify the minimum permissions an agent requires to do its job safely.

BUILDING MULTI-STEP & MULTI-AGENT WORKFLOWS

Topics Covered

• Workflow architectures: sequential chains, parallel execution, and the supervisor– subagent pattern
• Reusable Skills: packaging house style, report templates and standard operating procedures for repeated use
• Handoffs and state: passing structured output cleanly from one agent step to the next
• Automating recurring processes — scheduled reporting, inbox triage, and meeting-notes-to-action-items pipelines
• Knowing when not to build an agent: the case for a formula, a macro, or a five-line script

Learning Activities

Architecture walkthrough with whiteboard mapping; live demonstration of a two-agent handoff; team hands-on build; group presentation and critique.

Hands-On Exercise

Competitor Briefing Pipeline — Teams build a two-agent workflow. A Research Agent gathers and cites publicly available competitor pricing information; a Writer Agent consumes that structured output and produces a one-page briefing note in the organisation‘s house template. Teams then break the handoff deliberately to observe how downstream agents behave on malformed input.

Expected Outcome

Participants can architect and build a multi-step workflow in which agents pass work between each other reliably and can justify their choice of architecture against simpler alternatives.

GOVERNANCE, EVALUATION, RISK & COST CONTROL

Topics Covered

• Building an evaluation set: gold-standard inputs, expected outputs, and explicit pass/fail criteria
• Measuring accuracy, completeness and run-to-run consistency — why one good demo proves nothing
• Guardrails in practice: scope limits, prohibited actions, escalation triggers and audit logging
• PDPA 2010, confidentiality, data residency and internal AI usage policy for Malaysian organisations
• Token economics: cost per run, right sizing the model, context hygiene, and when Haiku beats Opus
• Change management: pilot design, user onboarding, and communicating agent limitations to stakeholders

Learning Activities

Presentation on evaluation methodology; demonstration of a scored evaluation run; individual exercise; case study of an AI governance failure; open Q&A on organisational policy.

Hands-On Exercise

Evaluation Rubric & Risk Register — Participants write a ten-case evaluation rubric for the workflow built in Module 6, execute it, score each case, and document the two weakest failure modes together with a concrete mitigation and the specific guardrail that would prevent recurrence.

Expected Outcome

Participants can prove — with evidence rather than impression — whether an agent is fit for deployment, and can present a defensible risk and cost position to management and IT.

CAPSTONE PROJECT — BUILD, TEST AND PRESENT AN AGENT

Topics Covered

• Scenario briefing and team formation (3–4 participants per team)
• Agent brief drafting, scope negotiation and success-criteria definition
• Build, test, refine — full application of Modules 1 through 7
• Team presentation, live demonstration, and structured peer and trainer critique

Learning Activities

Team-based case study; supervised hands- on build with trainer coaching at each table; live demonstration; assessed presentation. Hands-On Exercise Examining a full Business Scenario of a fast- moving consumer goods distributor company.

Expected Outcome

Participants demonstrate integrated command of the full agent lifecycle — from business problem to designed, built, evaluated and defended working agent.

Wrap-Up, Action Planning & Q&A

• Consolidation of key principles across both days and the agent design checklist
• Open Q&A on organisational rollout, licensing and IT approval pathways
• Personal action plan: three agents each participant will build within thirty days
• Recommended progression pathway — Claude Code, API-based agents, Power Automate integration, and AI governance policy development