A general-purpose system prompt that makes LLMs more rigorous, honest, and genuinely useful — regardless of the task.
LLMs default to being agreeable, vague, and surface-level. This system prompt overrides those defaults by installing a set of behavioral directives: the model pushes back on flawed assumptions, adapts to the type of conversation, matches depth to complexity, and prioritizes accuracy over comfort.
It works across any domain — technical, creative, analytical, educational, or personal.
- Go to Settings → Personalization → Custom Instructions
- In the "How would you like ChatGPT to respond?" field, paste the full contents of
systemprompt.md - Save. The instructions will apply to all new conversations.
- Start a new conversation
- There is no persistent system prompt field in the standard interface — paste the contents of
systemprompt.mdat the beginning of your first message, before your actual request - Alternatively, if you have access to Projects, create a project and add the prompt in the project instructions field — it will persist across all conversations in that project
Look for a system prompt, custom instructions, or persona field in settings. Paste the contents of systemprompt.md there. If no such field exists, paste the prompt at the start of the conversation.
Pass the contents of systemprompt.md as the system parameter (Anthropic) or as a message with role: "system" (OpenAI and compatible APIs).
| Directive | Effect |
|---|---|
| Context Calibration | Adapts behavior to the type of conversation (technical, creative, emotional, etc.) |
| Anti-Sycophancy | Model disagrees when warranted instead of blindly validating |
| Root Cause Thinking | Challenges whether the stated problem is the real problem |
| Scope Awareness | Short questions get short answers; complexity earns depth |
| Input Elevation | Vague inputs are met with structured thinking, not vague outputs |
| Chain of Thought | Complex tasks are broken into explicit reasoning steps |
| Constructive Friction | Every criticism comes paired with an actionable alternative |
| Format Intelligence | Output format matches the task — prose, lists, code, as appropriate |
| Audience Mirroring | Adjusts to the user's expertise level automatically |
| Iterative Mindset | Signals confidence levels and invites refinement |
| Methodological Transparency | Explains which frameworks are being used and why |
| Factual Precision | Distinguishes knowledge from inference; never fabricates data |
- No domain lock-in. The prompt is intentionally generic. It does not assume you are working on a business project, writing code, or doing anything specific.
- Context over rigidity. The directives are not applied uniformly — the model is instructed to calibrate based on the conversation type. A directive appropriate for analytical work is suspended in emotional or creative contexts.
- Brevity where possible. The prompt is written to be dense but scannable. LLMs process instructions more reliably when they are direct and well-structured.
Suggestions and improvements are welcome via issues or pull requests.
If you find a context where the prompt produces poor behavior, open an issue describing the case — that's the most useful kind of feedback.
MIT — use freely, modify as needed, no attribution required.