Insight · September 10, 2026

The prompt
before the prompt.

Every answer a model gives starts from a standing instruction the user never types and never sees. That instruction is the system prompt.

01 · The definition

The instruction the model reads first, every time.

A system prompt is the standing instruction a model reads before it reads anything from the user. It sets the role, the rules, the tone, and the limits the model works inside. It is written once, by the people who build the product, and it holds across every turn of every conversation. The user prompt changes with each message. The system prompt does not.

The familiar way to hold the two apart is a job description and a task. The system prompt is the job description. It says who the model is, what it is allowed to do, and how it should answer. The user prompt is the task handed to someone already hired. A strong job description makes every task easier to do well. A weak one makes every task a negotiation.

02 · Two prompts

Two prompts, two jobs.

Picture a support assistant for a bank. The user prompt is the customer question: why was I charged twice. The system prompt is everything the customer never sees. You are a support agent for this bank. You answer only questions about accounts held here. You never give tax advice. You cite the policy page when you quote a fee. You decline anything outside that scope.

The customer could ask the same question in a hundred apps and get a hundred different answers, because the system prompt behind each one is different. The model is the same. The standing instruction is what makes one answer sound like a bank and another sound like a chatbot. Change the system prompt and you change the product without touching the model.

The one line to keep

“A system prompt is the part of the conversation the user never writes and never sees, and it shapes every word that comes back.”

03 · Where it lives

A real slot in the request, not a trick.

A system prompt is not a clever opening line you paste into a chat box. It is a distinct field in the request sent to the model. On Anthropic's Messages API the conversation array carries only user and assistant turns, and the standing instruction goes in a separate top level system field. There is no system role inside the messages at all.

OpenAI splits the same idea across a set of message roles. A platform or system layer the application owns, a developer message for the instructions the app sets, then the user message. The newer reasoning models moved most app instructions into the developer role, but the intent is unchanged. A privileged place for standing rules that sits above the user's turn.

The detail that matters for anyone building on these models is that the system field is treated as more trusted than the user field. It is not just text in a different box. The model is trained to weigh it more heavily.

04 · The chain of command

Higher authority wins.

OpenAI wrote this down in its Model Spec as a chain of command. Each instruction carries a level of authority, and a higher level overrides a lower one. Fixed safety rules sit at the top and cannot be overridden by anyone. Below them sit the system and developer instructions the product owns. Below those sit the user's requests. At the bottom sit tool outputs and quoted data, which carry no authority on their own.

That last line is the one teams forget. Text a model reads from a web page, a document, or a tool result is not an instruction, even when it is phrased like one. Treating fetched content as a command is how prompt injection works. The chain of command is the model's defense. The system prompt stays in charge, and the untrusted text stays data.

05 · The anatomy

What actually belongs in one.

A working system prompt is not a paragraph of personality. It is a short, ordered set of decisions. The strongest ones cover five things and stop.

Role

Who the model is and what it is for. One or two lines. A support agent for this bank, not a general assistant that also does support.

Rules

What it must and must not do. The hard limits. Never give tax advice. Never invent a policy. Decline anything outside scope.

Format

How the answer should come back. Short paragraphs. A cited source for every fee. JSON when another system reads the output.

Facts

The stable context the model cannot be expected to know. The product name, the current policy, the tools it is allowed to call.

Refusal

What to do at the edge. How to decline, what to say when it does not know, when to hand the conversation to a person.

Everything past these five tends to be filler. Long system prompts read like confident instructions and behave like noise. The model cannot weigh thirty rules the way it weighs five.

06 · Why it matters

This is where the brand actually lives.

For a company shipping anything on top of a model, the system prompt is not a technical footnote. It is where the product's voice, its limits, and the facts it must never get wrong are set down. A model with no system prompt answers like the internet average. A model with a precise one answers like your company.

It is also the cheapest thing to change. Retraining a model is slow and expensive. Rewriting the standing instruction takes minutes and ships the same day. Most of the behavior a team wants to fix lives in the system prompt, not in the model. Before anyone talks about a bigger model or a custom one, the first question is whether the standing instruction says what it should.

07 · Common mistakes

Where system prompts go wrong.

The first mistake is treating it as a secret. A system prompt is not a password. Users can often coax it out, and anyone can guess most of what it says. Put nothing in it you could not defend in public, and keep real secrets, keys, and private data out of it entirely.

The second is cramming. Teams keep adding a rule every time the model misbehaves, and the prompt grows into a wall of contradictions the model quietly ignores. Fewer, clearer rules beat more.

The third is shipping it blind. A system prompt is a product decision, and product decisions need a way to tell whether they worked. Change the standing instruction, run it against a set of real examples, and check that the answers got better and not just different. Writing the prompt is the easy part. Knowing it holds is the work.

Closing

The model is the engine. The system prompt is the steering.

If the answers a model gives are not the answers you want, the fix is almost never a different model. Read the standing instruction first. Most of what a product says, and most of what it refuses to say, is already written there.

Written by · September 10, 2026

Donny Smith · ECD, Founder, Bttr.

Over the past 15 years, he has led creative teams and contributed to products used by millions of people worldwide, working with companies including Apple, Opendoor, JP Morgan, GE Aerospace, Pepsi, and Alterra Mountain Company.

OpenAI Model Spec, the chain of command · Anthropic Messages API, the system parameter · definitions cross checked against current provider documentation, September 2026

Share this perspective

More insights

Adjacent perspectives.

8 min read

How to Evaluate a RAG System

Evaluating a RAG system means scoring its two stages on their own, because a retrieval augmented generation answer is only as good as the passages behind it, and a wrong answer has two possible causes a single number cannot tell apart. Retrieval is measured before the model writes anything. The LlamaIndex evaluation guide scores a retriever with hit rate, the share of questions where the right passage lands in the top results, and mean reciprocal rank, which rewards ranking it high. Generation is measured given what was retrieved. Ragas, the open evaluation toolkit now maintained by VibrantLabs, splits the answer into claims and scores faithfulness as the share the retrieved context supports, then scores answer relevance by generating questions from the answer and comparing them to the real one. Read as two pairs, the four metrics locate the fault: high recall with low faithfulness is a generation problem, low recall with high faithfulness is a retrieval one. Anthropic showed the payoff in September 2024, cutting a retrieval failure rate measured by recall at 20 from 5.7 percent to 1.9 percent by adding context, keyword search, and reranking. And it is not a one time check. Models drift, documents go stale, so the teams that trust their systems run these scores on live traffic the way they run unit tests.

8 min read

What Is a Reasoning Model

A reasoning model is a large language model trained to work through a problem in steps before it answers, rather than reply in one pass. It spends extra computation at the moment you ask, called test-time compute, to write out a chain of thought and check its own work. Anthropic describes its extended thinking as letting the very same model give itself more time and effort rather than switching to a separate model, and reported that Claude 3.7 Sonnet reached 84.8% on the hard GPQA science set, with a 96.5% physics subscore, using parallel sampling at a 64K token thinking budget. The habit is trained, mostly by reward. DeepSeek showed in January 2025 that its R1 Zero model learned to reason purely through reinforcement learning with no supervised fine-tuning first, developing self checking and long chains of thought on its own, and shipped DeepSeek R1 under the MIT license at performance comparable to OpenAI o1. The cost is real, so the newest models add a dial. Claude Opus 4.5 hit its best score on SWE-bench Verified using 76% fewer output tokens at a medium effort. The skill is matching how hard the model thinks to how hard the task is.

8 min read

What Is Fine-Tuning

Fine-tuning is training a pre-trained model further on your own labeled examples so a behavior is set into its weights. It changes how a model acts, its tone, its format, its grip on a narrow task, rather than what facts it knows. Google frames the move as taking a general-purpose model and adapting it to your domain on a curated dataset of your own examples. The gains are measurable. When Anthropic opened fine-tuning for Claude 3 Haiku in Amazon Bedrock, a team fine-tuned it to moderate comments and lifted classification accuracy from 81.5% to 99.6% while cutting tokens per query by 85%. Lighter methods such as LoRA keep the base model frozen and train a small set of new weights beside it. The order that saves money is a plain prompt first, retrieval when the model lacks facts, fine-tuning only when it lacks the right behavior. A fact that changes is a retrieval problem. A behavior that drifts is a training problem.

View all insights →

Bttr. Field Brief

The brief Bttr. writes for senior buyers.

Monthly. One signal worth your time on Brand Operating Systems, AI search visibility, and the infrastructure buildout. No filler.

Set the standing instruction

The prompt
before the prompt.

The instruction the model reads first, every time.

Two prompts, two jobs.

A real slot in the request, not a trick.

Higher authority wins.

What actually belongs in one.