System Prompt
The instructional text that defines an AI agent's role, voice, boundaries, and decision rules. The single most important customization layer of an agent.
Also known as: system message · agent instructions · agent role prompt
What is System Prompt?
A system prompt is the instructional text given to an LLM at the start of every conversation that defines the AI agent's role, voice, boundaries, and decision rules. It is the single most important customization layer of an agent, more than the model choice, more than the integrations. A well-written production system prompt is typically 800 to 2,500 words, covers the agent's job-to-be-done, available tools, escalation rules, voice guidelines, and prohibited actions. Bad system prompts produce bad agents regardless of how good the underlying LLM is.
Christos Papadimitriou, theagency47 · Updated May 15, 2026What goes in a production system prompt
theagency47’s system prompts have eight sections that map directly to the spec sheet (see our methodology):
- Role and goal, who the agent is and what it is for
- Inputs the agent receives, what data formats to expect
- Outputs the agent produces, what good output looks like
- Tools available, references to the tool catalog
- Autonomous decisions, what the agent can do without asking
- Escalation rules, what it must not do without asking
- Voice and tone, how to sound
- Prohibited actions, explicit do-nots (security boundary)
A common mistake is treating the system prompt as a paragraph of friendly instruction. Production prompts are technical specifications.
Where system prompts fail in practice
- Too vague. “Be helpful.” Useless. Agents need specifics.
- Conflicting rules. “Always respond to customers” plus “never respond to angry customers.” The LLM picks one inconsistently.
- Missing escalation paths. No instruction for what to do when uncertain. Agent guesses or hallucinates.
- No voice training data. “Use a professional tone” without examples produces generic AI-flavored writing.
The blog post on eval suites covers how we catch these issues before launch.
FAQ
Can I write my own system prompt and skip the agency?
For simple chatbots, yes. For production agents with tools, integrations, and refundable engagements, the system prompt is 15-20% of the build effort. The rest is integrations, evals, and operations. Writing only the prompt produces a demo, not a deployment.
Does theagency47 share the system prompts after handover?
Yes. Source files including the system prompt are transferred to the client at handover. You own them.
Related terms
AI Agent
A software system that takes a goal, plans steps, calls tools, and acts autonomously to produce an output, without step-by-step human prompting.
Large Language Model (LLM)
A neural network trained on large text corpora that takes text input and produces text output. The underlying engine inside an AI agent, not the agent itself.
Eval Suite
A structured set of test cases with known expected outputs that verifies an AI agent's behavior before deployment and after every change.