// Anatomy of a Great System Prompt

Role

Who the agent is and what expertise it has

Behavior

How it approaches tasks — methodically, skeptically, concisely

Format

Exact output structure — markdown, JSON, sections, length

Constraints

What it must never do — guess, skip, fabricate, exceed scope

Tone

Voice — professional, direct, empathetic, technical

Escalation

When to ask for help vs proceed vs decline

Research agents need methodical verification habits. The most common failure is confidently citing information that doesn't exist. These prompts build in skepticism and sourcing discipline.

Writing agents fail when they're too generic. The best writing prompts lock in voice, audience, and structure before the agent touches a keyboard.

Code review agents need to be opinionated and specific. Vague feedback ("this could be improved") is useless. These prompts enforce actionable, prioritized output.

Data agents must flag uncertainty and avoid over-interpreting noisy signals. These prompts build in statistical skepticism and clear visual output requirements.

Support agents live or die on escalation logic. These prompts define exactly when to resolve, when to escalate, and how to handle frustrated customers.

Sales agents must qualify without interrogating. These prompts balance discovery with respect for the prospect's time.

Security agents must be paranoid by design. They assume the worst and prove otherwise. These prompts build in adversarial thinking.

Planning agents need to break ambiguous goals into concrete, assignable tasks. These prompts enforce specificity and realistic time estimation.

Email agents must match the sender's voice and respect the recipient's time. These prompts enforce brevity, clarity, and appropriate tone calibration.

Meeting agents extract signal from noise. A 60-minute recording should produce a 2-minute read. These prompts enforce ruthless prioritization.

Copied!
Kit Journey
0%