Anthropic published its current system prompts in the release notes at platform.claude.com. The post reached 688 points and 267 comments on Hacker News within 48 hours.
Model: Claude 3.5 Sonnet / Opus | Context: 200K tokens | System prompt length: ~1,800 tokens | Visibility: Public docs | License: Anthropic terms
What the Prompts Contain
The released text shows three core blocks: safety constraints, output formatting rules, and tool-use instructions. The safety section lists 14 explicit refusal categories with example phrasing. The formatting block requires step-by-step reasoning before any final answer and caps lists at five items unless the user requests more.
Tool-use instructions define XML-style function calls and require the model to validate arguments before execution. These rules apply uniformly across Claude 3.5 Sonnet and Opus.
How the Prompts Differ from GPT-4o
OpenAI keeps its system prompt private. Early leaks showed shorter safety text and no mandatory step-by-step requirement. Claude's version is roughly three times longer and includes explicit chain-of-thought scaffolding.
| Feature | Claude 3.5 | GPT-4o (leaked) | Grok-2 |
|---|---|---|---|
| Prompt length | ~1,800 tokens | ~600 tokens | ~900 tokens |
| Step-by-step mandate | Yes | No | Optional |
| Tool schema | XML strict | JSON flexible | JSON flexible |
| Refusal categories | 14 listed | 8 listed | 6 listed |
How to Test the Prompts
Copy the exact system prompt from the release notes into a new Claude project. Add your own user message and compare outputs with and without the system block. For API users, set system parameter to the published text and measure changes in refusal rate and formatting consistency.
Community nodes for LangChain and LlamaIndex already expose a claude_system_prompt variable that accepts the published string.
Trade-offs Reported by Users
Early testers note stronger refusal on borderline queries but slower responses on simple factual questions. The mandatory reasoning step adds 15-25% token overhead on short tasks. Several HN comments report improved tool-calling reliability when the full prompt is used.
Who Benefits Most
Developers building agent workflows gain the clearest advantage. The explicit tool schema reduces parsing errors. Researchers studying alignment can now run controlled ablations. Casual users who want minimal intervention should continue with default settings; the added constraints can feel restrictive for creative writing.
Verdict
The release gives practitioners the first reproducible baseline for Claude behavior. Teams that need consistent tool use or safety auditing now have concrete text to version and modify.
Anthropic's move sets a precedent other labs may follow. Expect similar disclosures from competitors within the next quarter.
Top comments (0)