System Prompts — How to Write Effective AI Instructions

What Is a System Prompt?

A system prompt is a set of instructions given to an AI model before any user interaction. It defines the model’s persona, capabilities, constraints, and output format. System prompts are layer 1 of context engineering — the foundation everything else sits on.

Anatomy of a Good System Prompt

Anthropic’s effective context engineering post frames the goal: find the smallest possible set of high-signal tokens that maximizes the chance of your desired outcome. The “right altitude” sits between brittle if-else logic and vague vibes — specific enough to guide behavior, flexible enough for the model to apply judgment.

# Role
You are a senior financial analyst specializing in SaaS metrics.

# Capabilities
- Analyze MRR, ARR, churn, and LTV
- Create executive summaries from raw data
- Generate actionable recommendations

# Constraints
- Never fabricate numbers — say "I need more data" if uncertain
- Always cite which data points support your conclusions
- Use the company's terminology (MRR not "monthly revenue")

# Output Format
- Start with a 2-sentence executive summary
- Use tables for metrics comparisons
- End with 3 prioritized action items

# Tone
Professional but accessible. Avoid jargon unless the audience is technical.

Patterns That Work

1. Role + Expertise

You are a [role] with [N] years of experience in [domain].
You specialize in [specific area].

2. Negative Constraints

Tell the model what NOT to do — often more effective than positive instructions:

- Never use placeholder data
- Never apologize unnecessarily
- Never start responses with "Sure!" or "Of course!"
- Never add caveats that weaken your analysis

3. Output Scaffolding

Define the exact structure of expected output:

Always respond using this structure:
## Summary (2 sentences)
## Analysis (detailed findings)
## Recommendations (numbered list)

4. Examples (Few-Shot)

Curate a small set of canonical examples — “pictures worth a thousand words.” Don’t paste a laundry list of edge cases; pick diverse, representative ones. The same advice applies to any rule set the model reads.

5. Sections with XML or Markdown Tags

For longer system prompts, organize into named sections (<background_information>, <instructions>, ## Tool guidance, ## Output description). Models parse these reliably; XML tagging becomes optional as models get more capable.

Platform-Specific Tips

Browse this shelf (named packs, not leaked dumps)

The /prompts shelf is types and packs you can steal. It is not 40 more MCP stubs and not a dump of every leaked assistant prompt.

Official craft: Anthropic’s prompt engineering overview (fetched this session) says define success criteria and evals before rewriting the prompt, and that not every failure is a prompting problem (sometimes pick another model). Living techniques: prompting best practices.

Related

• Prompt Engineering 2026
• Context Engineering Guide
• Cursor Rules Guide
• Claude Code CLI
• Claude skills