Why the 4-part prompt? An AI gives better answers when you stop "chatting" and start
giving it structure. Naming the Role, Context, Task, and Format
removes guesswork: the AI knows who to be, what it's working from, exactly what to do, and how
to shape the answer. It's the difference between a vague reply and a usable one — and it's
what readies the AI to work safely against a knowledge base instead of guessing.
Copy each section into your AI tool of choice — ChatGPT, Amazon Q, Claude, or any other.
Tip: the live demo answers only from the fictional "Commonwealth of Example" documents. Build a prompt about that agency (permits, policies, services) to see it work.
Example Prompts
Example 1 — Summarize from the Knowledge Base
RoleYou are a government policy analyst.
ContextUse the Commonwealth of Example IT Modernization Program documents.
TaskSummarize the program goals, current status, and the Permits project.
Format5 bullets, plain language, no acronyms.
Example 2 — Draft a Reply from the Service Docs
RoleYou are a constituent services representative for the Commonwealth of Example.
ContextA resident is frustrated about a delayed building permit. Use the modernization and service-catalog docs.
TaskDraft a reply that acknowledges the delay and gives a concrete next step.
FormatUnder 150 words, warm and professional, do not promise a specific date.
Example 3 — Explain an Incident from Jira
RoleYou are a platform engineer briefing a new teammate.
ContextUse Meridian Jira ticket MDS-1042 and the related architecture docs.
TaskExplain what happened, the root cause, and how it was resolved.
FormatA short summary, then root cause and resolution as bullets.
The 4-Part Prompt
Role→Context→Task→Format
Booster 1: Give an example (few-shot)
Booster 2: Iterate — tell it what NOT to do
5 Things to Try Monday
Summarize a long document (RFP, policy, report)
Draft a constituent reply with guardrails
Turn messy meeting notes into structured action items
Ask AI to explain a technical concept in plain language
Build a prompt template for a task you do every week
The "Never" List
Never put PII, CJIS, FTI, PHI, or secrets into AI
Never send AI output without human review
Never treat AI as a system of record
Never skip your agency's AI/data policy
Never assume AI output is correct — verify everything
How to Start
Get your knowledge searchable (organize docs, put them somewhere indexable)
Pick one repetitive task and try AI on it this week
Share what works with your team
AI in Action — Mass Digital Government Summit 2026 | Jamie Taschetti, Cloud Architect, SMX