System prompt
Persistent instructions given to an LLM before any user interaction — the frame within which the agent reasons.
Definition
The system prompt is a set of instructions loaded at the start of every LLM or coding agent session. It defines behavior, constraints, tone, and context in which the agent operates. In an agentic workflow, CLAUDE.md and AGENTS.md files serve as persistent system prompts — they're injected automatically at every session.
postcursors perspective
The system prompt is the most underrated tool in the agentic workflow. A well-written AGENTS.md file at the project root has more impact on agent response quality than the choice of model. It's the permanent memory that compensates for inter-session amnesia.
In practice
"Configuring the system prompt" for a coding agent means writing and maintaining the project's AGENTS.md file: tech stack, conventions, what the agent must not do without asking, useful commands.
Common misconceptions
- ✗ Putting project history in the system prompt — it should contain rules, not a timeline
- ✗ Ignoring the system prompt and manually briefing everything each session
The system prompt is to the coding agent what .editorconfig is to the editor: a configuration file that ensures consistency without manual intervention. The difference is that the system prompt influences reasoning, not just formatting.