← AI for Professionals: Using LLMs at Work
Lesson
Lesson 3: Capstone — build and improve a prompt for your own task
Build, critique, and improve a prompt for the learner's own real work task, applying role/context/task/format, constraints, iteration, verification, and privacy, checked against a rubric.
Bringing it all together
Bringing it all together
You have now covered the full arc: how LLMs work and why they hallucinate; prompt engineering moves (role, context, task, format, constraints, few-shot examples, chain-of-thought, structured output, iteration); how to apply AI to real work tasks; agents and RAG; and safety, privacy, and ethics. The capstone is your chance to transfer all of that to something you actually do at work.
A strong capstone prompt has five layers. (1) Role and audience: tell the model who it is and who will read the output — this shapes tone and depth. (2) Context and inputs: give the background the model needs; paste only safe, non-sensitive information. (3) Task and output format: state exactly what you want — a table, a list, a draft email, a JSON object — and what you do not want. (4) Constraints: length, language, what to rely on ('use only the information I provide'), what to avoid. (5) Verification plan: decide in advance what you will check — facts, numbers, legal claims, tone — and how you will check them before using the output.
Iteration is not failure; it is the method. After your first output, apply one targeted move: ask the model to critique its own answer, request a shorter version, add an example, or sharpen the role. Log what changed and why.
Privacy gate: before finalising your prompt, confirm — no credentials, no PII, no customer data, no trade secrets, no unreleased material — unless you are on an approved enterprise plan with the right agreements in place.
Lesson notes
Bringing it all together
You have now covered the full arc: how LLMs work and why they hallucinate; prompt engineering moves (role, context, task, format, constraints, few-shot examples, chain-of-thought, structured output, iteration); how to apply AI to real work tasks; agents and RAG; and safety, privacy, and ethics. The capstone is your chance to transfer all of that to something you actually do at work.
A strong capstone prompt has five layers. (1) Role and audience: tell the model who it is and who will read the output — this shapes tone and depth. (2) Context and inputs: give the background the model needs; paste only safe, non-sensitive information. (3) Task and output format: state exactly what you want — a table, a list, a draft email, a JSON object — and what you do not want. (4) Constraints: length, language, what to rely on ('use only the information I provide'), what to avoid. (5) Verification plan: decide in advance what you will check — facts, numbers, legal claims, tone — and how you will check them before using the output.
Iteration is not failure; it is the method. After your first output, apply one targeted move: ask the model to critique its own answer, request a shorter version, add an example, or sharpen the role. Log what changed and why.
Privacy gate: before finalising your prompt, confirm — no credentials, no PII, no customer data, no trade secrets, no unreleased material — unless you are on an approved enterprise plan with the right agreements in place.