Getting reliable results

Two prompts can ask for the same thing and get wildly different results: one three sentences, the other three pages; one a tidy list, the other a wall of prose. The model is not being unpredictable for no reason. It is filling in gaps you left open. Say the length and format you want, and most of that variation disappears.

Specifying length

Vague length words like "briefly" or "in detail" leave a lot of room. A model reads "briefly" differently depending on the topic. Concrete units work better: a word count, a sentence count, a number of bullet points, or a comparison such as "about the length of a tweet."

Specifying format

Format is separate from length. Do you want prose or bullets? A table? A numbered list you can follow step by step? Plain text or a code block? Name the shape you want the answer to take, and say it before the model starts writing, not after.

• Structure: bullets, numbered steps, a table, or plain paragraphs.
• Tone: formal, casual, technical, for a beginner.
• Output type: plain text, Markdown, JSON, or a code block in a specific language.
• Sections: ask for headers if you want a longer answer organized.

What temperature does to variety

AI Foundations covered temperature: a setting that controls how willing a model is to pick a less likely next token. Length and format instructions control the shape of an answer. Temperature controls how much the wording and choices vary between separate runs of the same prompt.

Most chat apps do not expose a temperature slider, but you can still nudge it with words. Asking for "one straightforward answer" pushes toward a low-temperature feel. Asking for "a few different creative options" pushes toward a high-temperature feel, even without touching a setting.

AI Foundations covered why models can be confidently wrong. This lesson is the practical follow-up: four habits that cut how often that happens, and a simple way to decide how hard to check.

Ground it in your source

When you ask a model about a document, a policy, or a piece of code, do not make it guess from memory. Paste the actual text into the prompt and ask it to answer using only that. This is called grounding. A grounded answer is built from what you gave it, not from a fuzzy recollection of something similar it saw during training.

Ask it to say when it is unsure

By default a model tends to produce a confident-sounding answer even when it is guessing. Asking directly for a signal changes that: "If you are not sure, say so" or "rate your confidence in this answer." It will not catch every mistake, but it surfaces some of the guesses instead of letting them read like settled facts.

Ask for sources

Asking "where does this come from" gives you something concrete to check: a source name, a document section, or a citation. AI Foundations already flagged that citations can be invented, so the point is not to trust the source blindly. The point is to have something specific enough to verify instead of a bare claim.

Check claims by stakes

You cannot verify everything, and you do not need to. Spend checking time in proportion to what happens if the answer is wrong. A brainstorm list costs little if it is off. A number that goes in a report, a medical or legal claim, or anything published under your name costs a lot. Sort by stakes first, then verify.

Every prompt eventually fails: the answer is off-topic, too generic, the wrong length, or just not what you meant. Before rewriting from scratch, run through a short diagnostic. Most bad results trace back to one of a handful of causes.

Working through the checklist

Start at the top. If the task is bundled ("summarize this and also check it for errors and also suggest a title"), split it into one ask, or number the parts explicitly. If the answer is generic, look for missing context: does the model actually have the source text, the audience, the constraint you have in your head but never typed?

If the content is right but the shape is wrong, that is a format problem, covered in the first lesson of this section: state the length and structure you want. If the model keeps missing a style or pattern you cannot easily describe in words, give it an example instead of more instructions.

Change one thing at a time

When a prompt is not working, it is tempting to rewrite the whole thing at once. Resist that. Change one variable, run it, and check the result. If you change the context, the format, and the wording all together and the new answer is better, you do not know which change did the work, and you will not be able to repeat it on the next prompt.