System Prompt Design
Set the agent's role, stable context, ground rules, capabilities, tools, and output format.
Prompt engineering is both a writing discipline and an operational discipline: define who the agent is, what it should do, the boundaries it must respect, and then manage those prompts like production code.
The deck divides the practice into six levers. The first four shape the prompt itself; the last two turn prompts into versioned, tested infrastructure.
Set the agent's role, stable context, ground rules, capabilities, tools, and output format.
Tell the agent precisely what to do, how to do it, and what successful completion means.
Define the lines the agent must not cross and back prompt rules with code-level enforcement.
Shape a consistent voice and character without allowing persona to replace real constraints.
Track immutable versions, diffs, promotions, audit history, and fast rollback.
Run each prompt version against a dataset, score outputs, compare with a baseline, then ship or iterate.
The system prompt is the agent's charter: stable, top-level context that frames every response.

Ambiguous instructions force the model to fill gaps with guesses. Good instructions reduce that space for interpretation.
“Summarize this document.”
“Summarize in exactly 3 bullets, each under 15 words, focused on financial risks. Output as a markdown list.”

Constraints define the boundaries the agent must not cross, regardless of what a user asks.

Persona is the voice and character users experience. It can improve consistency and brand fit, but it should never substitute for actual policy or controls.

Prompt craft determines the behavior you want. Prompt-Ops makes that behavior reproducible, promotable, observable, and reversible.
A prompt version captures more than words. It should reproduce the conditions that produced the behavior.

Agent outputs are non-deterministic, so testing is based on datasets, scorers, baselines, and regression checks rather than exact-string assertions.

Author, test, version, deploy, monitor, and improve. Production signals feed the next round of authoring.