Across reports, Playbooks, and chat, the quality of what Aline's AI gives back depends on how you ask. A vague prompt produces vague or inconsistent results; a precise one produces data you can trust.

This guide covers the prompt-writing techniques that make the biggest difference — and how to use the AI's own explanations to improve.

In this article you'll:

The AI does exactly what you ask, so ambiguity in the prompt becomes ambiguity in the output. Small wording changes — naming the exact clause, stating what counts and what doesn't — can be the difference between a column full of accurate answers and one full of false positives.

Name precisely what you're looking for. Instead of "find data agreements," specify the exact document types and how to tell them apart — for example, distinguishing a BAA from a DPA, and stating what each must contain to qualify.

💡 The more concrete the definition, the fewer false positives you'll get.

When you need a clause pulled word-for-word rather than summarized, include the keyword "verbatim" in the prompt. This is essential when the exact language matters — pulling a specific provision, a price-increase cap, or a defined term — and you can't have the AI paraphrasing.

Decide how the AI should respond when a clause isn't present. For completeness, it's often better to have it explicitly return "silent" or "not found" than to leave a blank — that way you can tell the difference between the AI didn't find it and the clause genuinely isn't there.

⚠️ If you're getting too many "not found" results, the prompt likely needs adjusting (see Step 5).

If a single prompt is trying to do too much, results get inconsistent. Break a complicated request into clearer instructions, and collapse long exclusion lists into a single condition where you can. Simpler prompts are more reliable and easier to debug.

Aline's AI explains its reasoning. When an answer is wrong, read the explanation to see why it answered the way it did — then adjust the prompt to close that gap.

This feedback loop is the fastest way to tune a prompt: change the wording, re-run on a small set, read the reasoning, repeat.

💡 Best Practice: You can iterate on prompt wording directly in AI chat before committing it to a report or Playbook. Try a phrasing, look at the output and the reasoning, refine, and only then apply it at scale.

Always validate a refined prompt on a small subset of documents you already know the answers to before running it across your full set. This confirms the prompt behaves as expected and saves tokens you'd otherwise spend re-running a large job.

Once it's accurate on the sample, run the full set with confidence.

Stuck on a prompt that keeps returning the wrong answer? Click the Support button in Aline and share the prompt plus one document it gets wrong — that's usually all we need to fix it.