One of the most common complaints from people who use AI every day: it's powerful, but unpredictable. You ask for short answers and get an essay. You ask for brevity and it turns robotic. You spend more time wording requests than doing the work. It's natural to conclude you're using it wrong.

Partly, you're not. Some of this is simply how these tools work. The rest is fixable, and the fixes are habits, not secret prompts.

Why it happens

It predicts, it doesn't obey. An AI assistant generates the most likely useful response to everything in front of it. Your instruction is one signal among many: the material you pasted, earlier messages, its own default style. When those pull in different directions, your instruction can lose.

Long chats dilute early instructions. Something you said twenty messages ago carries less weight than what's on screen now. In long conversations, assistants drift back to their defaults.

Vague rules get literal answers. "Be brief" doesn't say what to keep. So the assistant cuts the parts that made it sound human and keeps the bullet points. That's the robotic tone people complain about.

Answers vary by design. Ask the same question twice and you'll get different wording. That's normal, not a sign you broke something.

Eight habits for predictable output

1. Say what it's for and who reads it. "A summary for my manager, who has two minutes" gets a different result than "summarise this".

2. Describe the shape, not just the length. Instead of "be brief", try: "Two short paragraphs, plain language, no headings, no bullet points." Instead of "shorter", say what to drop.

3. Show one example. A sample of output you liked beats a paragraph of description. Paste it and say "match this style".

4. Save standing preferences once. Most assistants let you store instructions that apply to every chat or project (custom instructions, project instructions or similar). Put your tone, format and "never do this" rules there instead of repeating them.

5. Start a new chat when it drifts. If it keeps ignoring a rule, don't argue for ten messages. Open a fresh chat, put the rule at the top, and paste only what's needed.

6. One task per request. "Read this, summarise it, draft a reply and translate it" invites shortcuts. Do it in steps and check each one.

7. Ask what it's unsure about. End important requests with: "List any assumptions you made and anything you're not sure about." Hidden guesses become a checklist.

8. Check the parts that matter. Numbers, names, dates, quotes, and anything legal or medical: verify them yourself or ask for sources you can open. The rest you can skim.

Staying in control

The goal isn't to trust AI blindly or to recheck every word. It's to know which parts of the output carry risk, and to check those. That's also what keeps your own skills sharp: you're directing the work, not just accepting it.

If you're just starting out, read the companion guide first: where to start with AI: a four-week plan.