System Prompts

What a system prompt is

The instruction layer Claude reads before any user turn. It sets persona, rules, tools-of-trade, response style, and constraints. The user can't override it (in well-designed apps).

In the Messages API:

client.messages.create(
    model="claude-sonnet-4-6",
    system="You are an expert SQL reviewer...",
    messages=[...],
)

Or as content blocks for caching:

system=[
    {"type": "text", "text": "You are an expert..."},
    {"type": "text", "text": LARGE_KNOWLEDGE_BASE, "cache_control": {"type": "ephemeral"}},
]

Anatomy of a strong system prompt

Anthropic's recommended structure (for non-trivial agents):

1. Role / identity      — "You are a customer support agent for ACME..."
2. Core mission         — what success looks like, the single primary goal
3. Tone & style         — formal? terse? bilingual?
4. Capabilities & tools — what Claude can do, including tool names + purpose
5. Boundaries           — what Claude must NEVER do; escalation rules
6. Output format        — JSON schema, markdown, plain text
7. Examples (optional)  — few-shot showing edge cases

Principles vs conditionals

A common cert exam concept:

Principles ("Always cite sources") — apply universally, hold up under pressure.
Conditionals ("If user is angry, transfer to human") — fragile, easy to dilute.

Prefer principles. Stack too many conditionals and the prompt degrades — Claude starts missing rules. This is the "dilution" failure mode.

Dilution

A long system prompt with too many rules at the same priority level causes Claude to weight them poorly. Mitigations:

Move critical rules to the start AND end of the system prompt (recency + primacy).
Use a # CRITICAL header for must-never-violate rules.
Move stable reference material into a cached block and keep the rule-of-conduct prompt short.
Split into focused subagents instead of one mega-prompt.

System prompt vs user instruction