Giving good instructions

Most “the AI is bad” moments are actually instructions-were-vague moments. The difference between a useful response and a frustrating one is rarely the model. It’s how the request was phrased.

This chapter is a working set of habits that turn vague asks into specific ones. None of them are complicated. The first time you read this, you might think “obviously”. After a week of practice, you’ll catch yourself in old habits and notice the better answers your AI gives once you change them.

The shape of a good instruction

Three pieces, in any order: what, for whom or what context, with which constraints.

Vague: Write me an email.

Better: Write me an email to Mike accepting his kitchen quote, professional but warm, under 100 words, ending with a question about when work can start.

The second one lands without a back-and-forth. The first one will produce an email that may or may not be the email you wanted.

You don’t have to be this complete every time. As your AI learns your preferences (see Knowledge and memory), defaults get sharper. “Email Mike accepting the quote” eventually produces close to what you want because your AI knows your default tone, length, and sign-off. Until then, spending an extra sentence saves a revision.

Five small habits that pay back

1. Say who, when you mean someone. “Tell Mike I’ll be late” is unambiguous if there’s one Mike. If there are two (the contractor and your brother-in-law), you’ll get asked. Name the right one the first time.

2. Say what you’re working on, when it’s not obvious. Inside a project’s chat, you don’t need to. From the general chat or Telegram, you do. “On the kitchen project, what did Mike quote for the countertop install?” routes cleanly.

3. Give the bound before the body. “Under 150 words” or “three bullet points” or “in markdown, ready to paste” up front saves rework. Your AI will follow constraints stated up front more reliably than constraints added after the fact.

4. State the audience. “For my team”, “for a client who’s never seen the spec”, “for Lily, who hates jargon” all change tone and density. Audience is the single highest-leverage hint you can give.

5. Push back specifically. “Make it better” is a poor revision request. “Cut the second paragraph and rework the opening to start with the cost” is a good one. Specific revisions converge fast. Vague revisions waste turns.

What to say when you don’t know exactly what you want

Sometimes you have a vague sense and that’s all. That’s fine. Don’t pretend to specificity you don’t have. Instead:

• Ask for options. “Give me three different angles on this and I’ll pick one.”
• Ask for a draft you can react to. “Take a rough cut at this; I’ll tell you what to change.”
• Ask the AI to ask you. “What do you need to know to do this well?”

That third one is underused. Your AI is often better at noticing what’s missing than you are, and asking it directly is much faster than going back and forth on three half-formed attempts.

What to NOT say

A few patterns to retire:

• “Be creative.” It’s not a constraint. If you want a surprising angle, ask for an angle that’s surprising in some specific way. (“Surprising for someone who’s read a hundred announcements like this.”)
• “Be concise.” Useless without a target. “Under 100 words” is concise.
• “Make it sound natural.” Pick a writer you’d want to sound like, or describe the voice. “Like how I’d actually text Lily, not like a press release.”
• “Don’t be too AI-sounding.” It will try, but a positive instruction works better: “use contractions, no bullet lists, paragraph form, sound like a human who actually wrote this.”

Standing instructions vs. one-time asks

Every preference you state in a conversation is potentially something that should be remembered for next time.

If you find yourself saying “make it shorter” on every artifact, that’s a standing instruction. Tell your AI to remember it:

“Going forward, default to under 150 words on any draft email unless I specifically ask for a longer one.”

That goes into the knowledge graph as a standing rule. Your next ten emails are shorter without you having to ask each time.

Standing instructions are the single biggest source of “huh, it just knows what I want now” over the first month. Watch for repetition in your corrections; turn the repetition into a rule.

A short example, end-to-end

You: Draft a status update on the kitchen project for Lily. She wants the short version.

Your AI: [produces a 3-sentence draft]

You: Good shape. Cut the bit about the timeline; she already knows. Add a line about the dishwasher decision being still open.

Your AI: [revises]

You: Perfect. From now on, when I ask for an update for Lily, default to this length and skip timeline unless I ask for it.

Your AI: Got it. [REMEMBER: Status updates for Lily default to 3 sentences, no timeline, only mention timeline if explicitly asked.]

Three turns, one durable preference. Every future status update is closer to what you want without you having to spell it out again.

The patience curve

Working with your AI gets easier over time, not because the AI gets smarter (it doesn’t, model-by-model; you’re using the same one), but because the working relationship accumulates. The first week feels like talking to a competent stranger. By month three, it knows the cadence of your work, the people in your life, your defaults, your hot buttons.

The investment that pays the most: every time you correct something, take the extra five seconds to say “remember that for next time”. That five seconds compounds for the next year.