Service demand, procurement and staffing data sit in departmental silos, so queues and bottlenecks are discovered through complaints. Senior officers see where citizens are waiting and which approvals are stalled, and act before service targets are missed.
Service DemandPurchasing & ContractsStaffing & PostingsBuildings & UpkeepService Counters
A worked example for a typical Malaysian Federal / State Agency, not a description of any real company. Go through the six stages in order: paste each prompt into Claude, check the answer against how your business really works, then continue. Change the profile at the start to match yours.
Stage 1 of 6
Map how the operation really runs
Act as an operations consultant. I run a Federal / State Agency in Malaysia. Profile: A Malaysian federal or state agency delivering citizen-facing services through counters, portals and field offices — several thousand staff across states.
Interview me, one question at a time, to map how decisions actually get made today. Cover: the recurring meetings and reports, who decides what, which numbers people trust, and where decisions are late or made on gut feel.
Known friction: Easy access vs. Budget limits; Quick approvals vs. Proper process; Online vs. In-person service; Spending the budget vs. Spending it well; One standard vs. Local needs.
After 8 to 10 questions, summarise the five decisions that cost us the most when made late, in a table: decision, current owner, current delay, cost of delay.
Stage 2 of 6
Design the specialist agents
Using the decision map above, design 5 specialist AI agents for our Federal / State Agency: Service Demand, Purchasing, Staffing, Buildings, Counter Network.
For each agent give: the one question it answers every day, the data it reads, what it may recommend, what it must never do alone, and the format of its output.
Starting points for what each should watch:
- Service Demand: application volumes, queue times, channel shift to online, complaint themes
- Purchasing: open tenders, spend vs. allocation, supplier performance
- Staffing: vacancies, leave, training due, workload by office
- Buildings: maintenance backlog, downtime, energy use
- Counter Network: cost and demand per service point, mobile and digital alternatives
Keep every agent narrow. If an agent covers two jobs, split it.
Stage 3 of 6
Wire up data and compliance
For the agents above, list the data feeds needed. Likely systems: Case-management and queue systems; Financial and procurement systems; HR records (aggregated); Facility maintenance tickets.
Output a table per agent: field, source, how often it refreshes, data-quality risk, and what the agent should do if the feed is stale or missing. Mark every personal-data field and say how to minimise it.
Rules we operate under: Public procurement rules, Official Secrets Act, PDPA obligations (and the government's own data-sharing guidelines), and agency circulars. Recommendations only; decisions stay with accountable officers.
End with a short list of actions that always need human sign-off.
Stage 4 of 6
Design the morning brief
Design the single daily brief that pulls the 5 agents together for the Director-General, ready before the first meeting of the day.
Specify how the brief merges overlapping or conflicting recommendations, how it ranks items (impact, urgency, reversibility, confidence), and the exact fields per item: what to decide, recommended action, evidence, owner, deadline, and the cost of doing nothing.
Then write a realistic sample brief of five items. Items might resemble: Open two extra counters at the Kuantan office on Mondays — queues are 3x the norm / Extend the portal helpdesk hours until the e-filing deadline / Flag three tenders approaching the approval deadline.
Stage 5 of 6
Build and automate it
Turn the design into a working automation for our Federal / State Agency.
Needs: a scheduled daily run; the 5 agents run in parallel and each saves structured JSON; a final step reads them all and sends the brief by email or chat; every run is logged with inputs, outputs and cost.
Propose the build with Claude plus a workflow tool (for example n8n or the Claude Agent SDK). Give the steps in order, what happens when a data feed fails, and a human-approval step before anything that spends money or makes a customer commitment.
Stage 6 of 6
Roll out and measure
Write a rollout plan for the Federal / State Agency agent team that starts small and earns trust.
Stage 1: one agent, advisory only, run alongside the current process. Stage 2: add the other agents. Stage 3: switch on the daily brief. Stage 4: let agents prepare low-risk actions for approval.
For each stage give duration, owners, what "ready to move on" means, and the metrics to track: time to decision, how often recommendations are accepted, hours saved, and errors caught. Close with the three biggest risks and how we contain them.
