Google AI Studio Adds System Instruction Templates

AI Product Launches & Updates

• LaTeX & shortcuts update: Josh Woodward @joshwoodward announced improved LaTeX rendering (inline editing, copy/paste) and new keyboard shortcuts (Shift-Command-O for new chat, Shift-Command-K for search) in @GeminiApp (video, shortcuts).

• System instruction templates: Logan Kilpatrick @OfficialLoganK shipped a save & reuse feature for system instructions in @GoogleAIStudio, enhancing experiment reproducibility with Gemini.

• Persistent memory for agents: LangChainAI @LangChainAI announced cognee integration, adding persistent memory to LangGraph agents for context retention across sessions.

AI Tools & Applications

• Claude Skills system: Jason Zhou @jasonzhou1993 introduced Claude Skill, a modular skill format (prompt+tools/assets) to 10x Claude code, and launched a public repo for awesome agent skills (overview, repo).

• askplexbot migration: Aravind Srinivas @AravSrinivas recommended Perplexity Assistant users on WhatsApp switch to their improved askplexbot on Telegram for better performance.

• n8n × Gemini integration: Philipp Schmid @_philschmid published a getting started blog for n8n.io with Google DeepMind Gemini, detailing infrastructure setup for conversational automation.

Product Management Insights & Strategies

• LLMs as code ghosts: Philipp Schmid @_philschmid highlighted Andrej Karpathy’s analogy that “LLMs are ghosts of internet-scale code”, excelling at common patterns but struggling with novel implementations.

• AI bloat vs craft: Aakash Gupta @aakashg0 shared that Gemini’s Head of Product warned AI accelerates product bloat, making craft (elegant UX) the key differentiator over additional features.

• PM best practices framework: George Nurijanian @nurijanian recommended reviewing “what the best product managers do” by John Cutlefish, a monthly checklist for core PM principles.

AI Industry Developments & News

• Open datasets boom: Clement Delangue @ClementDelangue spotlighted trending open datasets—including Fineweb, Webscale-RL, and SVQ—arguing there’s no excuse not to train your own models now.

• On-device LoRA fine-tunes: Sebastian Raschka @rasbt observed that Apple’s models are LoRA fine-tunes running on-device, underscoring the commercial viability of edge adaptation.