Insights
Prompt Engineering Is the Smallest Part of the Work
Published 19 July 2026 · DataTranquil · 7 min read
Why isn't a good prompt enough to ship a reliable AI feature?
A prompt only controls what you ask the model to do; it has no control over what data it's working from, whether that data is accurate, or whether the output gets checked before a user sees it. Most of what determines reliability sits outside the prompt entirely, in the system built around it.
What is context engineering?
Context engineering is deciding exactly what information the model sees at the moment it's asked to answer — which records, which history, which retrieved documents — and in what order and format. A model with the right context and a mediocre prompt usually outperforms a perfect prompt fed the wrong information.
What is data engineering's role in AI system quality?
Data engineering determines what the model is actually working with underneath the context layer: whether records are deduplicated and current, whether fields mean what they claim to mean, whether the source of truth is even the source being queried. No prompt can compensate for a model reasoning over data that's stale or wrong.
What is eval engineering, and why does it matter more than the prompt?
Eval engineering builds the test set and scoring method that tells you whether the system actually works — on real cases, edge cases, and adversarial ones — before and after every change. It matters more than the prompt because it's the only part of the system that tells you the truth about failure rate.
So what should teams actually spend their time on?
Spend the least time on prompt wording and the most on the layers underneath it: the context pipeline, the data it draws from, and the evals that measure whether it works. Teams that treat the prompt as the whole project usually ship something that works in the demo and nowhere else.
Smallest effort
Prompt
Wording and instructions given to the model at call time.
Where reliability starts
Context
What records, history, and retrieved documents the model actually sees.
The foundation
Data
Whether the underlying records are accurate, current, and traceable.
The proof
Eval
The test set that tells you the real failure rate, before and after changes.
None of this means the prompt doesn’t matter — it does. It means treating the prompt as the whole project is the fastest way to ship something that impresses in a demo and fails the first time it meets data the demo never showed. The layers underneath are where the actual engineering happens.
Get started
Want to know where your own AI feature's effort should go?
An AI-readiness discovery reviews the context, data, and eval layers behind your prompt, not just the prompt itself.