System Prompts and Roles

The concept

Most modern AI tools let you set a "system prompt" — sometimes labeled "custom instructions," "a persona," or "a project configuration" — that persists across an entire conversation, rather than being just one more message in the chat. A regular message addresses one task at a time; a system prompt sets the standing frame for every message that follows: who the assistant should act as, what rules it should always follow, what format it should default to, and what it should never do.

The practical value is large. Instead of re-explaining "you're a concise technical editor, always respond in bullet points, flag unclear jargon" at the start of every message, you set that once as a system prompt, and every subsequent reply in that conversation — or, in tools with persistent custom instructions, every conversation — follows the same style automatically. This turns a general-purpose chatbot into something closer to a specialized tool built for one recurring job, without you having to re-brief it each time.

It helps to know where this setting lives in common tools, because the label changes even though the idea doesn't: ChatGPT has "Custom Instructions" in settings, which persist across all chats, and also supports a system-style first message when used through the API or in a "Project." Claude has "Projects" with custom instructions and a separate "Styles" feature for tone. Most AI features built into other software — browser extensions, writing apps, customer support tools — have an equivalent "assistant settings," "persona," or "behavior" panel, even when they don't use the words "system prompt" anywhere in the interface. Learning to recognize the pattern across tools is more useful than memorizing one tool's menu location.

A second thing worth understanding is scope. Some system prompts apply to a single conversation thread; others (like ChatGPT's Custom Instructions or a Claude Project's instructions) apply to every new conversation you start under that account or project. Confusing the two is a common mistake — people write a careful system prompt for one chat, assume it will carry over to their next unrelated question, and are confused when the tone reverts to generic. If you want a standing behavior across all your work, look specifically for the account-level or project-level setting, not just a one-off system message.

A third nuance: a system prompt sets a strong default, but a specific instruction in a regular message can still override it for that one exchange — for example, "ignore the bullet-point rule just for this answer, I need a full paragraph." System prompts are a frame, not an unbreakable law, but they dramatically reduce how much repetitive instruction-giving you need to do across a long working session.

Worked example

Picture setting this system prompt once, in a tool's custom instructions field: "You are a concise technical editor. Always respond in bullet points under 100 words. Flag unclear jargon in brackets." From that point forward in the conversation, every question gets a bulleted, sub-100-word answer with jargon flagged. Asking it to review a paragraph, summarize a document, or explain a concept all come back in exactly that same consistent style, without you ever repeating the instruction.

Compare that to a second, role-based system prompt for a different job: "You are a patient onboarding coach for new employees at a small logistics company. Explain things in plain language, assume no prior company knowledge, and always end with 'Next, you should...' pointing to one concrete next step." Under this system prompt, the same question — "how does our invoicing process work?" — produces a very different answer in tone and structure than it would under the technical-editor prompt above, even though the underlying model hasn't changed at all. The system prompt is doing all the work of shaping the output to the job.

Without a system prompt at all, you'd need to restate "keep it under 100 words, bullet points, flag jargon" (or the onboarding-coach framing) in every single message, and would likely forget to on at least one of them — producing an inconsistent conversation where some answers are terse and bulleted and others are long paragraphs, purely because you forgot to repeat yourself once.

Exercise

Set up a custom system prompt (or custom instructions/persona, depending on your tool) in an AI assistant for a role you'd use often in your actual job — for example, "resume editor," "meeting-notes summarizer," or "customer-reply drafter for a SaaS support team." Be specific about format and tone in the system prompt itself: name the exact output structure (bullets, a table, a fixed word count) and the voice (formal, casual, blunt).

Then run three different real requests through that same conversation or assistant and check: did the persona, tone, and format stay consistent across all three, without you repeating the instructions? If one of the three broke the pattern, note exactly what in your system prompt was ambiguous or missing — that's the part to tighten.

As a second pass, deliberately send one message that asks the assistant to break its own system prompt for just that reply (e.g., "for this one answer only, write a full paragraph instead of bullets"). Confirm it follows the override for that message, then check whether the very next message reverts back to the system prompt's default. This tells you concretely how the "strong default, not unbreakable law" behavior works in the tool you actually use.

Recap

• A system prompt is a sticky, standing instruction that frames an entire conversation (or every conversation, depending on scope), unlike a regular one-task message.
• Setting it once — with a specific role, format, and tone — turns a general-purpose assistant into a consistent, specialized tool for one recurring job.
• A specific in-conversation request can still override the system prompt for a single exchange; it's a strong default, not an unbreakable rule, so always verify scope and override behavior in the tool you use.

Resources

• Anthropic: "System prompts" — https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/system-prompts
• OpenAI Help Center: "Custom instructions for ChatGPT" — https://help.openai.com/en/articles/8096356-custom-instructions-for-chatgpt
• Anthropic: "Claude Projects" — https://www.anthropic.com/news/projects