Skip to content

Applied AIArticle

Prompting is not a strategy: designing reliable AI features

Better prompts help, but reliable AI features come from product design: narrowing the task, grounding the model, and planning for failure.

ActuateMinds EditorialPublished 2 min read

Where prompt tweaking stops paying off

Every team building with language models goes through a phase of prompt tweaking: add an instruction, rephrase a sentence, paste an example. It helps — up to a point. Past that point, reliability comes from the design of the feature, not the wording of the prompt.

1. Narrow the task

The broader the request, the more ways a model can go wrong. "Answer any customer question" is hard to make reliable. "Summarise this support ticket into three fields" is much easier.

Ask of every AI feature:

  • What is the smallest useful job the model can do here?
  • Can the job be split into steps, each with a checkable output?
  • Which parts can be handled by ordinary code instead?

2. Ground the model in your data

Models fill gaps with plausible text. Give them the facts they need, and tell them to use only those facts.

  • Retrieval: fetch relevant documents and include them in the request.
  • Structured context: pass the user's actual account data rather than describing it.
  • Citations: ask the model to point to the source passage it used, and show it to users.

3. Constrain the output

Free text is hard to validate. Structured output is easy.

{ "category": "billing", "urgency": "high", "summary": "Customer charged twice for March" }

Validate the output in code. If validation fails, retry once or fall back gracefully.

4. Design for being wrong

Assume the model will sometimes be wrong, and design the experience around that:

Risk Design response
Wrong but confident answer Show sources; let users verify
Harmful or off-policy output Filters, refusals and escalation paths
Irreversible actions Require confirmation before acting
Low confidence Hand off to a human or a standard flow

5. Measure in the product

Track the outcomes you care about: tasks completed, corrections made, escalations, user ratings. A feature that sounds impressive but is frequently edited or abandoned is not working.

A quick checklist

  • The task is narrow and its output checkable.
  • The model receives the facts it needs.
  • Output is structured and validated.
  • Failures are visible and recoverable.
  • There is an evaluation set and a launch bar.
  • Real usage is measured after launch.

Key takeaways

  • Prompts are one lever among many — and rarely the strongest.
  • Narrow tasks, grounded inputs and structured outputs drive reliability.
  • Design the experience assuming the model will sometimes be wrong.
  • Measure real outcomes, not impressions.

Written by

ActuateMinds Editorial

The ActuateMinds editorial team writes practical guides for people moving into product, design, data and applied-AI roles.

Get the next article in your inbox

One short email a week. No spam.

No spam. Unsubscribe from any issue.

Applied AI2 min read

Evaluating LLM features before you ship them

Demos are easy; reliable AI features are not. How to build a small evaluation set, choose criteria, and decide whether a feature is ready.

Industry Insights2 min read

What changes when every product role uses AI

Observations on how AI tools are reshaping the day-to-day work of PMs, designers and analysts — and which human skills are becoming more valuable, not less.

Product Management5 min read

How to write a PRD engineers actually read

Most product requirement documents are skimmed once and forgotten. A good PRD is short, opinionated about the problem and honest about what is still unknown. Here is a structure that earns attention from engineering.