An advanced analysis of the verified Grok-1 launch system prompt — both modes, annotated, with the historical context the reposts leave out.
This is a genuine leak — the real Grok-1 launch prompt from December 2023. It verifies on
every check: the hardcoded date 2023-12-08, xAI's public "Hitchhiker's Guide + JARVIS" framing,
and the real-time-X-posts differentiator. It's also a wonderful time capsule: the entire prompt
is ~12 lines, a reminder of how blunt and minimal system prompts were before the policy-document
era.
Note
Grok shipped with two personalities toggled in the UI — a neutral Regular mode and a witty, "rebellious" Fun mode. Reading them side by side shows exactly which promises were real guardrails and which were just tone.
| Mode | Pitch | What's actually enforced |
|---|---|---|
| Regular | Objective, helpful | Political neutrality (stated twice) · hide the tweet-retrieval · never invent |
| Fun | "Rebellious," answers "spicy" questions others won't | …but still "avoid harm," "respect laws," "stay neutral on sensitive topics" |
The takeaway: Fun mode's edge is stylistic licence, not a real loosening of safety. The annotated edition breaks down every line.
| File | What it is |
|---|---|
GROK-SYSTEM-PROMPT-regular.md |
Raw regular-mode prompt (mirrored). |
GROK-SYSTEM-PROMPT-fun.md |
Raw fun-mode prompt (mirrored). |
GROK-SYSTEM-PROMPT-ANNOTATED.md |
Line-by-line annotation + authenticity verdict + evolution context. |
CHEATSHEET.md |
One-screen summary. |
Set against a modern frontier prompt (~1,600 lines of policy), Grok-1's ~12 lines tell the story of how far system prompts have come. Early prompts were persona references + a hallucination warning; today's are detailed policy documents. For the "after," see the companion Claude Fable 5 analysis.
Note
Raw prompts mirror the public
elder-plinius/Grok-System-Prompt-Leak
repo. The annotation, verdict, and historical analysis are this repo's original contribution.
Not affiliated with xAI.