AI quick tips

The four-part prompt

Every prompt needs a goal. The other three parts are optional, but the more you add, the more usable the first answer will be.

Goal

What you want the AI to do. The only required part.

Context

Who it's for and why — shapes the tone and what counts as relevant.

Expectations

The format, length, and tone you want the answer back in.

Source

The specific file, email, or text the AI should work from — attach or paste it, don't just describe it.

Quick tips

• Be specific. “Summarize this” gets a vague summary; “summarize the pricing terms in under 150 words” gets a usable one.
• Give it the source instead of describing it. Attach the file or paste the email thread — don't retype what already exists.
• Always review the output. AI can get names, dates, and figures wrong — check anything before it goes to a customer or into a file.
• Before pasting customer or employee details into any AI tool, ask whether it actually needs the real name. “The customer” or a set of initials often works just as well until the final draft.

Picking a mode

Most AI tools let you trade speed for depth — usually a menu near the prompt box. The names differ between ChatGPT, Claude, Copilot, and Gemini, and the exact models change every few weeks, but the choices stay about the same. Your setting often sticks between chats, so check which one you're in.

Fast / standard

Answers right away with less reasoning. Good for a quick rewrite, a tone change, a subject line, or a short draft.

Thinking / reasoning

The tool takes longer and works through the problem before answering. Worth switching on for comparing documents, spotting inconsistencies, planning, or anything where it has to weigh several things at once.

Research / deep research

Runs many searches and reads many sources, then writes a cited report. Slow, but the right choice for vendor comparisons and open-ended questions. Check the citations — that's the point.

Watch for metered charges. Free and standard plans cap how much you can use the more powerful modes, and some agent features bill per use. Routine prompts won’t hit this, but check before running a big batch.

Tokens, context, and starting fresh

Every message you send, every reply, and every file you attach gets read together as one thing — the conversation’s context. It’s measured in small chunks called tokens (roughly a few letters or a short word each), and every chat has a limit on how many it can hold at once.

• Big attachments cost the most. A long contract or a stack of spreadsheets can use up a large share of the limit before you've typed a word, leaving less room for a careful answer.
• Long chats accumulate. The tenth question is answered with all nine before it still in view. Usually that's helpful, but old tangents, false starts, and earlier drafts are still “in the room” shaping the next answer.
• Answers can drift. Once a chat runs long or jumps between topics, the AI can start blending in details from earlier — the wrong customer name, a number from a different project, a tone left over from an earlier question.

When to start a new chat instead of continuing this one:

• You're switching to a different customer or project — a fresh chat keeps details from one out of the answer for the other, which matters for confidentiality as well as accuracy.
• The conversation has gotten long or wandered, and answers start feeling generic, slower, or slightly off.
• You're starting a genuinely new task, even for the same customer — a clean chat gives the AI a focused context instead of one crowded with an earlier task.

Don’t want to retype everything you’ve established so far? Ask the AI to write the opening prompt for you:

Worked example

Goal, context, expectations, and source, all in one natural prompt.

Two habits worth building in

• Tell the AI what not to do — don't add facts, don't guess numbers, don't change the tone or substance. That cuts down on overreaching.
• Ask it to point to where in the document it got something, so checking its work is a quick look rather than a re-read of the whole file.