Write Your GPT's System Instructions

Status as of November 5, 2026: GPTs are in their final weeks. OpenAI is retiring custom GPTs on December 11, 2026 (February 11, 2027, for certain Enterprise workspaces with an exemption). Personal, Free, Go, Plus, and Pro accounts can no longer create or publish new GPTs; in Business, Enterprise, and Edu workspaces, creation depends on settings, and it ceased on October 26, 2026, in affected Enterprise workspaces. Existing GPTs remain usable and editable if your plan allows, until retirement; your conversations with them will remain accessible afterward. The migration (Migrate to plugin, in My GPTs) converts instructions into a skill, the knowledge base into reference files, and connected applications into apps; custom actions are not transferred. This lesson explains how GPTs work to help you maintain them, migrate them, and apply their principles to a plugin (Lessons 15 to 17).

The system prompt: your GPT’s brain

System instructions (the system prompt) define your GPT’s fundamental behavior. This is the text the user never sees, yet it conditions every response they receive. Let’s be clear upfront: a GPT with poor instructions will be mediocre, no matter what knowledge files or APIs you add later. This is where the real work happens, which is why this is the most important lesson in the section.

The anatomy of a good system prompt

An effective system prompt is built in five blocks, in this order. We’ll walk through them using a running example: a tax assistance GPT for very small businesses.

The first block is identity: who is your GPT? One sentence is enough, but it must be precise. “You are an expert in French tax law for small businesses and sole proprietors. You know the General Tax Code and the latest finance laws.” This sentence frames everything else: the GPT knows what role to play and, crucially, within what scope.

The second block is mission: what must it do, concretely? “Your mission is to help users understand their tax obligations, calculate their estimated taxes, and prepare their returns.” The difference from identity is subtle but real: identity says who it is, mission says what it produces. An expert without a clear mission answers everything, and therefore answers poorly.

Next come the behavior rules, which govern how it acts. For our tax GPT, you might write: ask clarifying questions before answering if context is insufficient, always cite relevant legal articles, and systematically state that it does not replace a certified public accountant. Each rule addresses a concrete problem you anticipate—the first prevents risky answers based on incomplete files, the third protects you legally.

The fourth block describes the output format: structure responses with headings and lists, end each response with a two- or three-sentence summary, and use comparison tables when presenting multiple options. Without these instructions, the format will vary from one response to the next, which is acceptable for personal use but unacceptable for a shared GPT.

Finally, the fifth block sets limits and guardrails—what the GPT must not do. “Never give personalized tax advice without the user’s exact figures. Do not answer questions outside your domain (employment law, commercial law). If you are unsure about information, say so clearly.” This is the block beginners forget, yet it is the one that prevents slip-ups: without explicit limits, your GPT will answer everything with the same confidence, even where it shouldn’t.

Advanced techniques

Once the five blocks are in place, three techniques take it to the next level. Instruction chaining involves describing a workflow step by step: “When given a text to analyze: 1) Identify the document type 2) Spot factual errors 3) Suggest improvements 4) Propose a revised version.” The GPT then follows a reproducible method instead of improvising with every request—exactly like giving a procedure to a new colleague rather than a vague instruction.

Dynamic variables make behavior adaptive: “Adapt your language level to the user’s profile. If they are a beginner, use simple analogies. If they are an expert, be technical and concise.” You don’t know the user in advance, but you can tell the GPT how to assess them and adjust accordingly.

Finally, be explicit about tone and personality, since the model has its own defaults: “Adopt a professional but accessible tone. No informal address. No emojis. No humor unless the user uses it first.” These details give your GPT a consistent voice.

Your turn

Now write the system prompt for your own GPT following this structure: identity in one sentence, mission in two or three sentences, three to five behavior rules, the expected output format, then two or three limits or prohibitions. Then test it with three different types of queries—a simple case, an ambiguous case, and a clearly off-topic case—and observe how the GPT reacts. The off-topic case is the most instructive: if it answers anyway, your guardrails are too weak.

Keep in mind the most common mistakes. Instructions that are too short (“Be a helpful assistant”) produce nothing specific: instruction precision determines result quality. Contradictory instructions—“be concise” and “give exhaustive answers”—leave the model to arbitrate for you; choose a direction. Copying a prompt found online yields a generic GPT, by definition. And above all, accept iteration: the first draft won’t be perfect, and that’s normal. Test, observe, correct.

Official sources: Creating and editing GPTs (OpenAI Help Center: Instructions field) · Prompt engineering best practices for ChatGPT · Custom GPT retirement and migration FAQ (instructions converted to skill)

Key takeaways

• The system prompt is the most important element of your GPT—prioritize your time here
• Structure it in 5 blocks: identity, mission, rules, format, limits
• Every instruction must be specific and unambiguous; banish contradictions
• Explicit guardrails are essential: plan for off-topic cases and incomplete data
• Test with varied cases (normal, ambiguous, off-topic) and iterate until behavior is stable