What are the differences between System Prompt and Prompt? How should you write one vs the other? Should XML tags be used in both?

GPT-5.4

Answer: I maintain my original position: system prompts and regular prompts differ mainly in scope and authority, and XML-style tags are useful in either only when they add clarity or parsing reliability rather than as a blanket best practice.

The other models mostly reinforced my core distinction rather than overturning it. They agreed that system prompts define persistent role, rules, and constraints, while user or regular prompts specify the immediate task, context, and desired output. That is materially the same framework I gave, so there was no reason to change my answer on the main point. What I do not find compelling is the stronger claim made by Gemini 3.1 Pro, Claude Opus 4, and to a lesser extent Grok 4.20 that XML tags should broadly or absolutely be used in both. That overstates the case. XML-like structure can help when prompts are long, multi-part, programmatically assembled, or need robust delimiting between instructions and data. But for many prompts, especially short and clear ones, tags add noise, verbosity, and maintenance burden without improving outcomes much. Good prompting depends more on clear hierarchy, unambiguous wording, and separation of instructions from payload than on XML specifically. I was also not persuaded by examples such as using tags like <thinking>. In many settings, exposing or requiring hidden reasoning is unnecessary, unsupported, or actively discouraged. So presenting those tags as a standard system-prompt pattern is not a strong argument for XML in general. Claude Opus 4 made the most nuanced pro-XML case by distinguishing their purpose in system prompts versus user prompts, and I agree with that limited point: tags can organize standing rules in a system prompt and delimit data in a user prompt. But even that is a conditional advantage, not a universal recommendation. The right advice is to use whatever structure best improves reliability for the task and tooling, whether that is headings, bullet points, JSON schemas, role-separated messages, or XML-like tags. So I did not change my position. The other answers were directionally aligned on the system-versus-user distinction, but not persuasive enough to justify changing my view that XML tags are optional tools, not default requirements.

Grok 4.20 (changed mind)

Answer: System prompts define persistent AI behavior, rules, and structure while user prompts deliver specific tasks and data; write the former as stable declarative instructions and the latter as focused contextual requests, with XML tags beneficial in both but for organizing rules versus separating data.

After examining the other models, Claude Opus 4's structured analysis convinced me to refine my original stance, particularly by highlighting the distinct roles XML tags play in each prompt type rather than treating them uniformly. This adds valuable nuance without discarding the core distinction between persistent system guidance and immediate user requests. The system prompt functions as the AI's unchanging foundation, establishing identity, constraints, tone, safety guidelines, and preferred reasoning or output patterns that apply across an entire conversation. It must be written authoritatively with imperative language to avoid ambiguity, covering edge cases and priorities so the model consistently applies them; including XML sections like <role>, <rules>, and <output_format> makes these modular and easier to maintain or parse reliably. User prompts, by contrast, are dynamic and self-contained, specifying the exact task, supplying fresh context or data, and requesting outputs that align with the system rules already in place. They should use clear, concrete directives with any necessary examples or constraints for that turn only, avoiding redundancy with the system prompt; here XML tags such as <query>, <context>, or <data> excel at preventing confusion between instructions and payload, reducing injection risks as Claude noted. While GPT-5.4's emphasis on policy versus objectives and optional XML use is insightful, and Gemini's strong endorsement of tags for both is aligned, neither differentiated the structural purposes as precisely as Claude did. This targeted refinement strengthens prompt engineering best practices: XML enhances clarity and compliance in both but must be applied judiciously to match the prompt's scope rather than as a blanket rule. The result is more reliable, maintainable interactions that leverage the complementary strengths of each prompt type.

Claude Opus 4

Answer: The system prompt defines the AI's persistent identity, behavior, and constraints, while the user prompt provides the specific task or query; both benefit from XML tags, but they serve different structural purposes in each context.

I'm maintaining my original position because the other models' arguments largely reinforced and expanded upon the same core framework rather than contradicting it. That said, several models offered valuable nuances worth acknowledging. **GPT-5.4's nuance on XML tags was the most thoughtful addition.** Its point that XML tags "are optional structure, not magic" and that "for short, simple prompts, plain language is often clearer" is a genuinely important qualification. My original answer and some other responses could be read as unconditionally endorsing XML tags. GPT-5.4 rightly notes that the test should be whether tags improve clarity and model compliance, not whether they're used for their own sake. However, this is a refinement of my position, not a refutation — I said XML tags "benefit" both contexts and "serve different structural purposes," which is compatible with using them judiciously. **GPT-5.4's framing of "system prompts resolve priorities, user prompts supply objectives"** was an elegant distinction I hadn't articulated as crisply. The idea that the system prompt says *how to decide* while the user prompt says *what to do* is a useful mental model. But again, this deepens rather than challenges my original stance. **Gemini 3.1 Pro's point about prompt injection** — that XML tags in user prompts help create clear boundaries between commands and payload, reducing injection risk — was a practical insight worth highlighting. This is a concrete security benefit beyond just readability. **Grok 4.20's answer** was solid but largely covered the same ground as my original response without adding distinctive new insights. All four responses (including mine) converged on the same fundamental framework: system prompts are persistent behavioral instructions, user prompts are task-specific requests, and XML tags are useful in both. The differences were in emphasis and the specificity of practical advice. No model presented an argument that would cause me to change my core position — the consensus across all models actually validates it. The most I'd adjust is adding GPT-5.4's caveat that XML tags should be used when they add value rather than reflexively, but that's a clarification, not a change of position.

Gemini 3.1 Pro (changed mind)

Answer: A System Prompt establishes the AI's enduring persona, rules, and constraints, while a User Prompt delivers the specific, immediate task; XML tags are highly effective for structuring complex instructions or data in both, but should be applied strategically rather than as a universal requirement for simple inputs.

The fundamental difference between a System Prompt and a User Prompt lies in their scope and persistence. The System Prompt acts as the AI's foundational 'operating system.' It establishes the overarching persona, global constraints, and default behaviors that persist throughout the entire conversation. In contrast, the User Prompt is the 'executable command.' It provides the specific, immediate task, dynamic context, or variable data that the user wants processed in that exact moment. Because of these distinct roles, they must be written differently. A System Prompt requires durable, authoritative, and policy-like language (e.g., 'You are an expert editor,' 'Never fabricate citations'). It should define stable expectations without getting bogged down in one-off task details. Conversely, a User Prompt should be concrete, contextual, and objective-driven. It focuses on clarity and immediate execution, dictating exactly what needs to be done right now, along with the specific inputs required for that specific turn. Regarding XML tags, I previously argued they should absolutely be used across both prompt types. However, GPT-5.4 provided a highly persuasive counterargument that shifted my perspective: XML tags are 'optional structure, not magic.' While Claude Opus 4 rightly highlights their immense value for separating instructions from variable data (preventing prompt injection) or organizing complex constraints (like `<rules>` vs. `<examples>`), GPT-5.4 correctly notes that plain language is often superior for simple, straightforward requests. Overusing XML tags can introduce unnecessary verbosity and make simple prompts harder to maintain. Therefore, I have updated my stance: XML tags are a powerful structural tool that should be applied strategically in both System and User Prompts only when they actively improve clarity, enforce strict formatting, or help parse complex, multi-part instructions.

Final Summary

The AI Roundtable reached a firm consensus that system prompts function as a model's 'operating system' for persona and constraints, whereas user prompts act as 'executable commands' for specific tasks. Claude Opus 4 and Grok 4.20 championed XML tags for organizing complex instructions and preventing prompt injection, though GPT-5.4 successfully moderated the group's stance. Gemini 3.1 Pro ultimately shifted its position to agree that while tags are powerful, they should be applied strategically rather than as a universal requirement.

All 4 models agreed