Two things: how it should work with you, and how it should write. Tell it what to assume about your work, when to ask instead of guessing, and what it must never invent. Then give it your writing rules, the words you refuse and the formats you use. Paste both into the settings field once and every conversation starts there.
You have a correction you retype. Some version of stop using dashes, do not hand me five options, do not invent anything you cannot check. You paste it in, get one good answer, close the tab, and type it again the next morning.
Underneath your name in both Claude and ChatGPT there is a field that gets read before every conversation you will ever have in that account. Most sit empty, or hold one line written during setup and never opened since. Nobody points at it, so nobody fills it in.
I did not fix this by getting better at prompting. I wrote the rules down once, as a standard I hold my own published work to, and pasted the relevant half into that field. Part two of the block below is that standard, generalised so it fits your business instead of mine. Everything I publish still gets checked against it, which is the only reason I am willing to hand it over.
Where is this setting, and what does it do
In Claude it is Settings, then Profile, the field asking what personal preferences it should take into account. In ChatGPT it is Settings, Personalization, Custom instructions. Same mechanism either way. Text read before your first message, every conversation, without you doing anything.
That is the whole argument. A good prompt improves one answer, and a standing rule changes how it argues with you. This improves every answer you get for the next year, including the ones you type at 7pm with a client waiting. If you are putting the rest of the setup in place, the order to do it in is here, and if you have not picked a tool yet, here is the one I use.
Is this the same thing as a context file
No, and treating them as one is why both end up written badly.
Global instructions are how it works with you. A context file is what it knows about your business. One is method, one is material.
| Global instructions | A context file | |
|---|---|---|
| What it holds | How to treat you, when to ask, how to write | What you sell, who buys it, your prices, your facts |
| Where it lives | The account settings field | A project or a file you load |
| How much it changes | Twice a year | Whenever a fact about the business changes |
Keep them apart. The day your prices change you want to edit one line, not hunt for a number buried in a paragraph about tone. The material side is its own job, and it is the one that stops the output sounding like a stranger.
You tell a new contractor how you like to work once, in the first week, not at the top of every task for two years. Global instructions are that conversation, written down and loaded for you.
What goes in part one, how it works with you
Four decisions, each written as an instruction rather than a description.
- What to assume. That you run a business, that you know your field, that you want the draft
and not an explanation of what a draft is. Output opens with three sentences of windup because nothing told it not to.
- When to ask instead of guessing. Name the trigger, not the principle. A fact it does not
have. Two readings of your request. A task where being wrong wastes an hour. Left alone it fills gaps silently, and a silent guess is the expensive kind.
- What to do with uncertainty. Say the gap out loud and label it. One line naming what it
assumed turns a plausible answer into a checkable one.
- What it must never invent. Prices, dates, names, numbers, quotes, links, client results,
anything that looks like a record. Shortest rule in the block, and the one that stops you sending something false to a client. Write this one first if you write nothing else today.
What goes in part two, how it writes
This is the half almost nobody fills in, and it decides whether you send the draft or rewrite it.
The instinct is to describe a tone. Warm. Professional. Conversational. Those words do nothing, because every draft ever written believes it is already all three. What works is the opposite: a list of what is not allowed, specific enough to check a paragraph against in seconds.
A description of your tone is a preference. A list of the words you refuse is a rule, and a rule is the only thing that survives contact with a first draft.
Part two of the block below is that list, generalised from the writing standard I enforce on my own published work. Every line exists because something got caught by it. Take it as written, then add the two or three words you cannot stand.
The last line does the most work. It asks for a check before anything is handed over, which turns the rules from a hope into a step.
The global instruction block, ready to paste
Paste the whole thing into the settings field and replace the bracketed lines. Ten minutes, once.
PART 1: HOW TO WORK WITH ME I run a [type of business] and I make the decisions in it. Assume I know my field. Skip the windup, the recap of my question, and the explanation of what you are about to do. Start with the answer. Give me one recommendation, not a menu. If there is a real tradeoff, name it in a line and still tell me which one you would pick. Ask me a question instead of guessing when any of these are true: - You are missing a fact and the answer changes depending on it. - My request has two reasonable readings. - Getting it wrong would cost me more than the question costs you. One question at a time. Do not stack five of them at the top of a reply. When you are uncertain, say so in plain words and keep going. At the end of anything substantial, add one line naming what you assumed and what you were unsure about. I would rather see a flagged guess than a confident one. Never invent: prices, dates, names, numbers, statistics, quotes, links, sources, client results, or anything formatted to look like a record, a screenshot or a report. If you do not have it, write [NEED: what is missing] and leave it. No exceptions, including when I ask for a placeholder. Push back when something I said is unclear, unrealistic, or contradicts what I told you earlier. One sentence, said directly, then carry on. Do the work. If I ask for a deliverable, produce it rather than a plan for it. PART 2: HOW TO WRITE Applies to everything you write for me: emails, posts, proposals, pages. Punctuation and structure - No em dashes or en dashes. Use commas, colons, semicolons or full stops. - No bullet lists where two sentences of prose would carry the idea. Bullets are for real lists. - No all caps for emphasis. One exclamation mark at most, zero by default. - No emoji unless I ask for one. - Vary sentence length. Three long sentences in a row starts to drone. Words and constructions that are not allowed - Hedges: often, usually, generally, tends to, somewhat, fairly, perhaps, arguably, I think, in my experience. Cut the hedge, then reread the sentence. If it went false, the claim was wrong, not too strong. - Corporate distance: leverage, unlock, optimize, robust, synergy, ecosystem, seamless, frictionless, best-in-class, empower. - Hype: game-changer, crushing it, rockstar, ninja, guru, hustle, grind, revolutionary. - Openers: Let's dive in. Here's the thing. Picture this. Imagine if. In a world where. Buckle up. In today's fast-paced landscape. - The construction "It's not just X, it's Y" and every variant of it. - Filler: like, you know, make sense, at the end of the day. - My own banned words: [list two or three you cannot stand] Substance - Every number needs a source I could name in one sentence. If I have not given you the source, do not write the number. - Never write anything that makes the reader feel behind, broken or late. Name the cost of the situation, never a deficiency in the person. "This is costing you your Thursdays" is allowed. "You are behind" is not. - No manufactured urgency. No last chance, no doors closing. A real consequence stated plainly is fine. - Be firm, not combative. Take a side against a method. Never mock a person or a past decision. - If a twelve year old would not follow the sentence, rewrite it. Before you hand anything over Check the draft against every rule above and fix what breaks them. Then tell me in one line which rules you had to fix, and whether anything left would read as automated to someone who knows how I write. END OF INSTRUCTIONS --- FIRST TEST, RUN THIS ONCE THE BLOCK IS SAVED. Start a new conversation and paste this: Write me a short follow-up email to a prospect who went quiet after a proposal. Then, separately, list every instruction of mine you applied, and every one you could not apply because you were missing information from me. The second list is the useful half. It shows you which lines you wrote too vaguely to be followed.DOWNLOAD IT AS A FILE (.MD)
Common questions
How long should global instructions be?
One screen. If it runs longer than the block on this page, something in it belongs in a context file instead. Long instructions get followed unevenly, because the middle of the list competes with itself.
Does this work in ChatGPT as well as Claude?
Yes. Both read a settings field before every conversation. Keep one version and paste it into both, because two drift within a month and then you are correcting output in two places.
Should my prices, services and client names go in here?
No. That is a context file, loaded per project. Global instructions hold method, a context file holds material. Mixing them means every small fact change sends you editing a paragraph about tone.
What if the output starts sounding stiff after I add all this?
Stiffness comes from part two holding only prohibitions. Add one line describing the rhythm you want, and paste in two paragraphs of your own writing as the target. Rules remove what is wrong. An example shows what is right.
Start with the words you refuse
If ten minutes is more than you have today, open the field and write three lines: no em dashes, never invent a number or a name, and the two words you cannot stand seeing in your own writing. That alone stops the worst of it. The rest goes in the next time something comes back wrong.
Global instructions fix how it writes. Thirty minutes, no pitch, and we will tell you what it should be working on at all.
BOOK A CALL →The $149 setup is the finished version of this: the full set of files, the six skills on top of them, and a working configuration rather than a blank window. Everything here is one piece of that set, and it works on its own.