System prompt
The initial instruction text given to an LLM that sets its persona, tools, constraints, and default behavior for the session.
Every production agent in 2026 ships with a carefully tuned system prompt — often 1–4k tokens — that defines who the agent is, what tools it can use, what it can refuse, and how it should format outputs.
The system prompt is the single highest-leverage knob in the stack. Most "this agent isn't working" problems are system-prompt problems, not model problems.
Public leaks of system prompts (Claude, GPT, Devin) routinely run to 10k+ tokens with explicit examples and edge-case handling. Production system prompts are real software.
Frequently asked
Can users see an agent's system prompt?+
Most agents try to keep them confidential, but skilled prompt injection routinely extracts them. Don't put secrets in system prompts; treat them as eventually-public.