4 [AINews] The Two Sides of OpenClaw
4d ago · 1 source confirmed single-source
Latent Space newsletter covers Claude Opus 4.7 launch, computer-use agents, and open-source agent research developments.
Anthropic announced Claude Opus 4.7 and Claude Design, a research-preview tool for generating prototypes and design assets, with benchmarks showing competitive performance against Gemini and GPT-5.4. The newsletter highlights emerging patterns in agent reliability, where simple harnesses and scaffolding outperform larger models, and covers research on agent self-improvement, web skills extraction, and open-world evaluations. Open-source agent stacks like Hermes continue proliferating, with new work on monitoring reasoning degradation and local inference systems enabling practical agentic workflows.
unprecedented levels of security incidents (60x more reports than curl, at least 20% of skill contributions malicious)
2 Changes in the system prompt between Claude Opus 4.6 and 4.7
3d ago · 1 source confirmed single-source
Simon Willison analyzes system prompt changes between Claude Opus 4.6 and 4.7, documenting updates to child safety, tool integrations, and response behavior.
Anthropic released Claude Opus 4.7 on April 16, 2026, with notable system prompt updates documented by Simon Willison. Key changes include expanded child safety instructions, new integrations with Claude in Chrome, Excel, and PowerPoint, and modifications to encourage more concise responses and tool-assisted problem-solving. The analysis reveals Anthropic's approach to model behavior refinement through system prompt evolution, with removals of outdated guardrails reflecting improved model capabilities.
Anthropic are the only major AI lab to publish the system prompts for their user-facing chat systems.
2 Claude system prompts as a git timeline
3d ago · 1 source confirmed single-source
Simon Willison documents Claude system prompt changes across versions using git timeline visualization.
Simon Willison created a git-based timeline visualization of Claude system prompts by converting Anthropic's published prompt documentation into separate versioned files. He used Claude Code to automate the conversion and assigned fake commit dates to enable browsing changes through GitHub's interface. The project enables detailed analysis of prompt evolution, such as differences between Claude Opus 4.6 and 4.7.
Anthropic publish the system prompts for Claude chat and make that page available as Markdown. I had Claude Code turn that page into separate files for each model and model family with fake git commit dates to enable browsing the changes via the GitHub commit view.
1 Adding a new content type to my blog-to-newsletter tool
4d ago · 1 source confirmed single-source
Simon Willison used Claude Code to add beat content support to his blog-to-newsletter tool via a single agentic prompt.
Simon Willison describes how he used Claude Code to extend his blog-to-newsletter tool to include a new "beats" content type (releases, tools, museums, and other external content). He provided Claude with a concise prompt that instructed it to clone his blog repository for reference, modify the HTML app to filter beats with descriptions, and validate the changes using local testing. The agent successfully generated the exact SQL and JavaScript changes needed, including a UNION clause and beat-type display mappings.
Coding agents can clone code from GitHub, and the best way to explain a problem is often to have them look at relevant code.