Custom Instructions: The Highest-Leverage Five Minutes in AI

Marc Andreessen posted his own custom AI instructions back in May, and the shape of it stuck with me more than any single line in it: a written contract for how the model should behave, right down to “never hallucinate or make anything up.” A separate tip changed how I write the actual requests: ask Claude to interview you first, because most of what goes wrong in a prompt is context sitting in my head that never made it onto the page. Andreessen’s instructions fix what comes back. The interview habit fixes what goes in. Those two things became the sheet I actually use.

After I wrote both halves down, my efficiency and effectiveness with Claude went up sharply. It’s the highest-leverage thing I’ve done with my own Claude, and it’s the first thing I now point people to, whatever their AI client of choice, before a project, before a connector, before anything else. Last week’s post covers what a project folder or a live connection to your inbox or Drive adds once the instructions are already in place. This is what comes first.

Most people skip it and go straight to connecting the AI to their inbox or shared drive instead, on the assumption that more access means better answers. Five minutes with a plain text file gets you further than that connection ever will, and it costs nothing.

The System

A good instruction sheet is short enough to read in one breath, and every line in it is something you’d say to a new hire on day one, not a wish list.

The first line is the floor, and it’s close to verbatim what’s on mine: “never hallucinate. Verify facts, figures, citations, names, and dates. If you don’t know, say so.” It borrows two words from Andreessen’s post, and those two words are exactly what critics went after hardest: you can’t instruct a model out of hallucinating just by telling it not to. They’re not wrong. But the sentence doesn’t stop there. You can’t order a model into being correct, but you can order it to check what’s checkable and admit what isn’t, and that’s the part doing the actual work. It’s the line that catches a model’s default failure mode, the confident paragraph filling in for something it doesn’t actually have. Instead of guessing at a number or a fact and presenting it as settled, it says so and asks for what’s real.

The second line changes how it argues with you, not just what it makes up: “Don’t flatter, hedge, or soften to be agreeable. If I’m about to do something dumb, say so plainly.” Skip this and most models default to agreeable. They’ll nod along with a bad plan as readily as a good one, because agreeing costs nothing and pushing back risks sounding unhelpful.

The third line buys you a chance to catch a mistake before it’s made instead of after: “Before executing on anything non-trivial, briefly state the plan so I can course-correct.” A one-line “here’s what I’m about to do” costs five seconds to read and saves the fifteen minutes of redoing it once it went the wrong direction.

Shrunk down, that’s don’t invent, don’t flatter, plan first, three beats you could say out loud in a meeting without opening the doc.

Then there’s the part that decides how many questions it asks before it starts. Match the depth of questioning to the stakes. A typo fix or a quick rewrite, it should just do it. A client-facing document or a decision with real money behind it, it should ask until it actually understands what you want before it starts. The test worth writing down is: would a good mentor execute now, or ask one more thing first?

The most useful part of the sheet, and the part most people never think to write down, is that you can tell it which mode to run in.

Left alone, it runs in Default mode: match the questions to the stakes, state the plan first, the mentor test above.

Say “just do it” or “JDI” and it flips into Just Do It mode. Zero clarifying questions, no plan first, it goes with its best read of what you meant. If it had to guess at something non-obvious, it lists the guesses at the end, not the start, so you’re not reading a paragraph of caveats before you get to the actual answer. Reach for this when the ask is low-stakes and you’d rather clean up an assumption than wait through a round of questions.

Say “interview me” and it goes the other direction. It doesn’t try to answer or build anything yet. It asks one focused question at a time and waits for your answer before the next, and it pushes on the assumptions you’re not saying out loud. It’s closer to being coached through a problem than handed an answer, and it’s the mode for the messy, half-formed version of an ask, before you’ve decided what you’re actually asking for.

Same Sheet, Three Modes

Picture someone new to Claude asking it to draft a note to a client explaining why a deliverable is running a few days late, but now the sheet above is sitting in front of it.

Ask without naming a mode and it runs Default: it asks what actually caused the delay and what the new date really is instead of inventing a plausible excuse, then states the plan, a short, direct note, before it writes anything. It reads more like a draft than a finished email, and that’s not a downgrade. A draft that admits what it doesn’t know yet is something you can act on. A polished note built on a guessed date is the one that goes out to the client and comes back to bite you.

Say “just do it” instead, because a two-line delay note you’re going to edit yourself doesn’t need five follow-up questions, and it writes straight through, flagging its guess about the reason for the delay in a short line at the end instead of stopping to ask.

Say “interview me” before a harder version of the same problem, like how to reset expectations with a client who’s already frustrated, and it stops trying to write anything at all. It asks what’s already been communicated, how frustrated the client actually is, whether the new date is real or another guess, one question at a time, until the two of you agree on what the note needs to do.

Nothing about the model changed across any of those three runs. What changed was which mode you told it to run in, and it followed the instruction.

Try It Yourself

Start smaller than the full sheet below. Write one line, the verify-and-say-so rule, and run it for a day. You’ll feel where it’s thin fast, usually on stakes: it asks three questions about a quick rewrite, or it barrels through something that actually mattered without asking anything.

Add the don’t-flatter line next and the plan-first line after that, each for a day or two of real use, then layer in the three modes once those three are second nature. That’s the order mine came together in, and it’s the fastest way to end up with a sheet you actually keep using instead of one you paste in once and forget.

Writing this from scratch takes a few passes. The first draft is usually either too vague to do anything (“be helpful and accurate”) or long enough that nobody reads past line three. Mine took several rounds of actually using it before it stopped needing edits, and you’ll know yours has landed when you stop catching a bad assumption after the plan already ran instead of before it started.

Paste a version like the one below into Settings under Instructions for Claude, or Personalization for ChatGPT, and it applies to everything you ask from then on. Anything specific to one project or one client belongs in that project’s own instructions instead, not the account-wide sheet.

The sheet above is deliberately the small version, three rules and three modes, because that’s the version people actually keep using. The full instructions library has more of both, for whenever three rules stop covering what you’re running into. Five minutes with either one still beats the next connector.