How to Tune Your AI's Parameters (Without Guessing)

Nobody teaches this part. Prompt-writing gets all the attention, and the settings panel next to the text box gets ignored until it's actively in your way. That's a real gap: the panel is on every major AI product, it changes the answer you get, and picking badly is one of the most common — and least visible — ways people get worse results from AI than they should.

The three dials every major AI app has

Strip away the branding and ChatGPT, Claude, Gemini, and the rest all expose some version of the same three controls.

Model — quick vs. smart. Every provider ships at least one fast, cheap, “good enough for most things” model alongside a slower, more capable one. OpenAI calls its version Instant vs. Thinking. Anthropic's Claude lets you pick the model tier directly. Google's Gemini splits into Flash and Pro. The names differ; the trade-off doesn't. The quick model is faster and, on a paid plan, cheaper per answer. The smart model reasons more carefully and handles ambiguity, multi-step logic, and nuance better.

Effort — low, medium, or high. Separate from which model you're using, most products let you dial up how hard it thinks before answering — sometimes called reasoning effort, sometimes “extended thinking.” Low effort answers fast, in roughly the time it takes to read the question. High effort can take noticeably longer because the model is working through the problem in steps before it commits to an answer, closer to a research assistant than a search bar. Treat high effort as the everyday stand-in for a full “deep research” mode — it gets you most of the benefit without the multi-minute wait.

Search — on or off. This one is binary: can the AI look things up on the live web before answering, or is it working only from what it already knows? Off is faster and keeps the answer focused on what you gave it. On is essential the moment the answer depends on anything that could have changed since the model was last trained — prices, news, someone's current job title, this morning's inbox.

A rule of thumb that covers most cases

You don't need a flowchart. Three questions get you there:

Could you answer this yourself in under a minute, or is the answer already sitting in the text you pasted in? Use the quick model and the lowest effort.

Is getting it wrong actually expensive — in money, time, relationships, or your job? Turn effort up, even if the answer takes longer to arrive.

Does the answer depend on anything that could have changed recently? Turn search on. If it doesn't, leave it off — it doesn't just cost time, it can pull in information you didn't ask for and didn't need.

Six everyday tasks, correctly configured

The mistake nobody warns you about

Most advice about AI settings only covers one direction: don't under-power a hard task. The other direction matters just as much and gets skipped almost everywhere.

Turning on high effort and web search for a one-line email isn't just slower — it can make the answer worse. An over-thought response to a simple ask tends to read as stiff, over-hedged, or padded with caveats nobody needed. On a paid plan you're also spending real money to get an answer you could have had in two seconds. Matching the setup down to the task is as much a skill as matching it up.

FAQ