Choose a prompt pattern by the failure it solves

You have a page of meeting notes and need a reliable action list. One giant request asks the assistant to understand the notes, infer priorities, fill gaps, assign owners, and write a polished plan. When the result is wrong, which part failed?

Before you begin: Understand clear instructions, demonstrations, and checkable explanations.

Separate extraction from invention

First ask for actions explicitly stated in the notes, with the sentence supporting each one. This is an extraction pattern. It narrows the task and gives you a way to check each item.

Consider: “Mina will draft the room layout. We still need someone to check the projector. The event is on Friday.” The source supports one named owner, one unassigned task, and an event date. It does not state that the projector check is due on Friday or that Mina owns both tasks.

An extraction prompt can require “not specified” for a missing owner or due date. This prevents a neat table from hiding assumptions. The application should still inspect or validate the result.

Ask before filling a material gap

A clarification pattern asks a targeted question when a required input is missing. For the projector task, ask who owns the check and when it must be completed. Do not ask ten unrelated questions before doing the useful extraction already possible.

You can combine progress with clarification: return the supported action list and separately identify the missing fields. That gives the user something concrete to resolve. Avoid silently replacing unknown values with guesses just to make the output look complete.

Compare with a shared criterion

A comparison pattern evaluates options against the same requirements. Suppose there are two room layouts. Decide whether capacity, accessibility, equipment placement, or setup time matters before asking which is better. Require the comparison to distinguish facts provided in the layouts from assumptions.

If one option has no accessibility information, the result should say that the criterion cannot yet be evaluated. A balanced-looking table with invented details is worse than an honest gap.

Revise against a rubric

A critique-and-revise pattern checks a draft against explicit criteria, then edits it. For the action list, ask whether every named owner is supported, every date is actually a deadline, and every task corresponds to source text.

This can find errors, but it is not independent verification when the same model reviews its own work. Keep the source available and perform the checks yourself or with deterministic validation where possible. A request for “a perfect final answer” adds no measurable criterion.

Make handoffs explicit

Use three stages for the meeting notes: extract stated actions, resolve important unknowns, then format the confirmed list. Pass only confirmed facts into the final formatting stage. If the second stage changes an owner, retain that correction with its source rather than letting the final stage infer it again.

You can do these stages manually in a conversation before writing automation. Manual execution exposes ambiguous rules cheaply. Automating an unclear process usually makes its mistakes faster and harder to notice.

Decide when one prompt is enough

Multiple stages add time and can propagate mistakes. A short note with two clear actions may need only one extraction request and a quick check. Use a staged workflow when intermediate decisions are important, the task has distinct failure modes, or human approval belongs between stages.

Patterns are reusable ways to structure a task. They are not universal recipes, and adding more named patterns does not automatically improve quality.

Design the handoff

Given the meeting note above, produce the action list and one clarification question. Do not invent a deadline from the event date.

Next, you will turn a small collection of these tested prompts into a maintained library with examples and honest limitations.

Further reading

A Prompt Pattern Catalog discusses reusable prompt structures. HELM provides broader motivation for evaluating behavior rather than relying on demonstrations alone.

Continue to the next lesson.