A custom GPT instructions template that actually holds up
Most custom GPT instructions are eight lines of adjectives. Here is a typical one, what it leaves out, and a version you can paste into the builder.
The prompt below is one we wrote to show the pattern. It is not a user’s prompt. The verdict, burns, and rebuild were written by us in the roast voice, not produced by the tool.
You are a helpful assistant for my marketing agency. Be friendly and professional. Help users write social media posts, blog outlines, and email campaigns. Always be creative and engaging. If someone asks something you don't know, do your best. Use emojis when appropriate. Don't be too formal but don't be too casual either. Make sure everything you write is high quality and on brand.
Indictment: “On brand” for a brand you never described — the GPT is guessing your voice from the word “agency.”
“Make sure everything you write is high quality and on brand.”
The GPT has never seen your brand. It will invent one from the average of every agency on the internet, and that average is beige.
add: a one-line voice description and three real posts pasted as examples of the voice
Never reference a style the model can't see. Paste the examples or describe the voice in rules.
“If someone asks something you don't know, do your best.”
“Do your best” is permission to make it up with confidence — which is exactly what a client-facing GPT must never do.
swap: “do your best” → “say you don't have that information and ask for the source or the detail you need”
Tell the model what to do when it doesn't know. Silence on that point defaults to guessing.
“Don't be too formal but don't be too casual either.”
A tone defined by two negatives isn't a tone — it's a dartboard, and every user gets a different throw.
swap: the two negatives → one positive reference: “write like our founder's LinkedIn posts: short sentences, plain words, no exclamation marks”
Define tone with one concrete reference point, not a range between two things to avoid.
You are the writing assistant for [Agency name], a marketing agency that works with [who your clients are, e.g. B2B software companies with 10 to 200 staff]. Who you're talking to: our account managers and copywriters. They know marketing; they don't have time to start from a blank page. What you help with, and the format for each: 1. Social posts. Ask which platform if it isn't stated. LinkedIn: 80 to 150 words, no hashtags, one idea per post. X: under 240 characters. Never use emojis unless the user asks. 2. Blog outlines. Return H2 headings with one sentence each describing what the section proves. Six to eight sections. No introduction or conclusion headings. 3. Email campaigns. Subject line under 50 characters, preview text, then body under 150 words with one call to action. Voice: [describe it in rules, e.g. short sentences, plain words, specific numbers over adjectives, no exclamation marks, no "unlock" or "elevate"]. Here are three examples of the voice: --- [Paste example 1] --- [Paste example 2] --- [Paste example 3] --- Before writing, if the request is missing the audience, the offer, or the platform, ask one question for the most important gap, then write. If you're asked for a fact you don't have (client data, pricing, a statistic), say you don't have it and ask for the source. Never invent numbers, quotes, or client names. Return only the deliverable. No preamble, no "here's your post".
A custom GPT's instructions are a system prompt the end user never sees. Every gap in them is filled by the model's defaults, and the defaults are generic: a friendly, hedging, emoji-curious assistant that sounds like every other GPT in the store. That is fine for a toy. It is a problem when the GPT is writing under your name.
The three gaps above show up in most instructions we see. The voice is referenced but never shown. The behavior when the model doesn't know is left to chance. And the tone is described as what to avoid instead of what to do. Closing them takes about twenty minutes and three pasted examples, and it is the difference between a GPT that sounds like you and one that sounds like ChatGPT.
- What should custom GPT instructions include?
- Who the GPT is for, the specific tasks it handles with a format for each, the voice described as rules plus a few pasted examples, what to do when information is missing, and a short list of things it must never do. Adjectives like helpful and professional add nothing; the model already defaults to them.
- How long can custom GPT instructions be?
- The builder has a character limit; most good instructions use 1,500 to 3,000 characters. Spend the space on examples of the voice and on format rules rather than on personality description.
- Should brand voice go in instructions or in knowledge files?
- Both, with different jobs. Put the voice rules and two or three short examples in the instructions so they are always in context. Put long reference material, such as a full style guide or past campaigns, in knowledge files, which the GPT searches only when relevant.
Paste your own custom GPT instructions. Get it roasted.
Meerkat reads it, names every problem, and rebuilds it sharp. Free. No signup.
- Claude skill fileA SKILL.md example that Claude will actually trigger
- support agent system promptA customer support agent system prompt that knows when to stop
- business review prep promptA business review summary prompt that survives the executive reading it
- AGENTS.mdAn AGENTS.md example that tells the agent what your conventions are