Before you build an AI agent, write its instructions as a standard: one job, written down the way you would brief a capable temp. Cover the job, the steps, the tone, the boundaries, and what done looks like. Then start with the smallest tool that works and keep a person on anything that sends.
You tried an AI tool. The output was odd, and you quietly decided the problem was you. It was not. Nobody told you the writing comes first. The tools get the attention; the instructions decide the result.
I build agents for a living, and every one starts as a document, not a tool. I call the method Standard First. Below: what agent instructions are, how to write them, one full example to copy, and what still breaks. If the word "agent" itself feels fuzzy, start with what AI agents actually are, in plain English, then come back.
On this page: What instructions are · Why agents fail · Write the standard · The six parts · A full example · Where it goes · What breaks · In short · FAQ
AI agent instructions are the standing rules an AI agent works from every time it runs: the job, the steps, the tone, the boundaries, and what a finished result looks like. A prompt is one request. Instructions are the rules that govern every request after it.
Tools give the same document different names: instructions, custom instructions, project instructions, or a system prompt. This method calls the version you write first the standard: your rules for one job, written where a machine can follow them. Write it once, in plain English, and it works in any of those boxes. The document is portable; the tools are not.
Most AI agents fail for a boring reason: nobody wrote the job down. The tool got picked first, the instructions arrived in fragments through a chat box, and the agent filled every gap with guesses. It did what it was told. It was told very little.
Writing the job down sits on nobody's list, and that is the whole problem. It gets paid for anyway: evenings of prompt-tinkering, drafts in a voice that is not yours, a subscription that quietly lapses. The fix is not a better tool. It is half a page of writing, done before any tool gets opened.
THE ORDER HOW MOST PEOPLE MEET AGENTS PICK A TOOL TINKER WITH PROMPTS ODD OUTPUT QUIETLY GIVE UP STANDARD FIRST PICK ONE JOB WRITE THE STANDARD SMALLEST TOOL THAT WORKS TEST AGAINST THE STANDARD Same tools, opposite order. The top path is how most people meet AI agents. The bottom path is the whole method.An agent standard is one job written down the way you would brief a person: what comes in, what good looks like, what to do with anything unusual, and when to stop and ask. Six moves, in order:
Six parts, and the whole thing fits on a page:
Here is a complete agent standard, shown whole. The job is drafting a launch email in my voice, the exact first job my building guide recommends. I wrote it fresh for this article; there is no client data in it.
AGENT STANDARD: LAUNCH EMAIL DRAFTS · LAST REVIEWED JULY 2026
THE JOB. Draft the email that announces something new to my list. One email per request. Drafts only: nothing is ever sent.
WHAT COMES IN. What we are launching, in my words. The date it goes live. One link. The last launch email, as the example to sound like.
THE STEPS. 1. Read the example email first. 2. Write a subject line of seven words or fewer. 3. Open with the news in one sentence. 4. Two short paragraphs on what it does for the reader. 5. Close with the one link and an invitation to reply.
THE TONE. Write like: "We built this because you kept asking. It is live today." Never like: "We are thrilled to announce an exciting new solution." Short sentences. Warm. No exclamation marks.
THE BOUNDARIES. Never send anything. Never invent a price, a date, a feature or a quote. Never add a discount. Use only facts given in the inputs.
DONE, AND WHEN TO ASK. Done means one draft, under two hundred words, that passes the tone lines above and uses only the facts supplied. Anything missing or unclear comes back as a question, not a guess.
Notice what it is not: code. It is a page of decisions. Every line narrows what the agent may do or defines what good means. Anyone who knows the job can write one.
You do not need special software to use an agent standard: the same document works pasted into a chat, saved as project instructions, or loaded as a system prompt. One check first: some jobs need no AI at all. Invoice reminders are a built-in setting in apps like Xero and QuickBooks. For everything else, the ladder runs like this, and the first rung works today.
Rung one: paste and ask. Open a free chat app, paste the standard, paste the inputs, ask for the draft. Nothing is connected, nothing can reach a customer, and it costs nothing.
Rung two: save it once. Claude calls the saved home a Project, ChatGPT calls it custom instructions, Microsoft's Copilot has its own version. The names move around; the document does not. Names checked July 2026, so glance at the menu before you rely on mine.
Rung three: wire it into a workflow, so the inputs arrive on their own and the drafts land where you work.
Rung four: an agent with a human gate. The agent works from the standard, the drafts queue up, and you press send. Nothing goes out on its own.
One aside for the technically minded: instruction files with names like AGENTS.md are the same idea living inside a code project. The building guide below covers that world.
Costs, in one honest paragraph: the writing costs your time, once. The first rung is free. Paid plans are priced like ordinary software subscriptions, not salaries, and a professional build is a one-off fee on top. The dear part is the writing, and you can do it yourself.
When the standard exists and you want the full build, the next read is how I build AI agents with Claude, step by step. It starts from this exact document.
Most agent problems trace back to the brief: the agent did what the instructions said, and the instructions did not say enough. A standard shrinks that gap. Nothing closes it. Four things still go wrong.
It follows what you wrote, not what you meant. "Keep it short" means nothing to a reader with no common sense to fall back on. Say how short.
Judgement drifts without examples. Rules alone cannot carry tone. The two real examples in the tone section do more work than everything around them.
Confident nonsense still happens. An AI can state something false and sound completely sure. So anything that matters gets read by a person before it goes anywhere.
Standards rot. Prices change, wording changes, the job quietly shifts. Put a date on the document and reread it quarterly. It is the shortest quarterly review you will ever chair.
Honest aside: if you would rather write the standard and have someone else build the agent, that is the exact shape of what I build for people. Or keep reading and do it yourself, because you genuinely can.
Most people call it prompt engineering, though for standing instructions it is closer to writing a clear job description. A prompt is one request; instructions are the rules an agent works from every time it runs. The tools keep changing, and the writing is the skill that lasts.
What is the difference between agent instructions and skills?Agent instructions are the standing rules for who the agent is and how its one job gets done. A skill is a saved pack of know-how an agent loads when one kind of task comes up. The instructions are always on; a skill gets picked up when its task appears.
Is an agent standard the same as a system prompt?No, though they are close relatives. The standard is the document a person writes: the job, the steps, the tone, the boundaries. A system prompt is one of the places that document can be loaded so a model follows it. Write the standard first and the system prompt is mostly written; the build guide shows the mechanics.
Is there a template for AI agent instructions?A useful template for AI agent instructions has six parts: the job, what comes in, the steps, the tone, the boundaries, and done plus when to ask. A generator cannot fill those in, because it has never seen the job or the business. Write each part in plain English, the way a capable temp would need it.
Is an agent standard just an SOP?Nearly. An agent standard is a standard operating procedure written for a reader that takes every word literally. So it adds what a person would fill in from common sense: tone examples, boundaries, and a rule for when to stop and ask. An existing SOP is a head start, not the finished document.
The pages behind the claims above. All were checked and loading on 18 July 2026.
I build agents for a living, always standard first, with a human gate anywhere a mistake would cost money or trust. If you have one job in mind and you would rather hand the build over, the entry point is one agent doing one job, built on your standard. Email me at ameera@asurekha.com and describe the job in plain words. If you have written the standard, attach it. That is the whole brief.