AI Compass

News

What’s new, with source and context.

• 27 October 2026Agent Skills: Anthropic hands the skill format to the Linux Foundation
• 27 October 2026AI as a document checker: run a counter-test and a repeat test before relying on it
• 27 October 2026Claude Design moves into Claude: claude.ai/design closes on 14 December
• 27 October 2026Claude in October: Sonnet 4.5 retires on 30 November, Haiku 5.5 is new
• 27 October 2026Cloudflare AI Gateway: Auto Router picks the AI model for each request
• 27 October 2026Cloudflare Workers AI: AI responses contained other customers' data

• 27 October 2026DIGICAP: medical device dossiers are meant to become machine-readable
• 27 October 2026FDA agenda for 2027: final AI lifecycle guidance planned, but no date set
• 27 October 2026GPT-6.1 Terra: OpenAI model at $1 input and $5 output per million tokens
• 27 October 2026n8n 2.44 is stable: prompt caching for Claude has to be switched on per node
• 27 October 2026UK: AI medical devices to be monitored across their whole lifetime
• 27 October 2026UK government confirms indefinite recognition of CE-marked medical devices
• 2 October 2026AI Act for medical devices delayed to August 2028: more time, same work
• 2 October 2026AI in regulatory work: grounded in sources, not a text generator
• 2 October 2026Generative AI medical devices: what FDA has put up for discussion
• 2 October 2026FDA: radiology AI still needs a 510(k), including follow-on devices
• 2 October 2026New standards for AI: EN 18286 and ISO/TS 24971-2 are out, presumption of conformity is not

Knowledge cards

How to work with AI agents, step by step.

• Checks the agent cannot bypasscheckpoints and gates

  Automation

• Does a skill still help?measuring skills

  Prompting

• "Done" only with evidencea check command instead of self-reporting

  Automation

• Keeping context cleanfewer tokens, clearer answers

  Costs

• Knowledge base for agentsthe LLM wiki after Karpathy

  Prompting

• Long runs and parallel workClaude Code's built-in tools

  Automation

• Maintaining your Claude Code setupclear out after every model change

  Prompting

• Memory in three tierswhat Claude knows about you and your projects

  Prompting

• Project map instead of rereadingClaude Code with a code map

  Costs

• Project start with rules and plan files

  Prompting

• Self-improving loop with locked checks

  Automation

• Separate building and checkinga critic with fresh context

  Automation

• Vet third-party skills before they run

  Prompting

• Decision modelswhen a model should only choose

  Automation

• Have your own website checked from the outside

  Automation

• Websites without the AI lookdirections, references, checks

  Website

• A second opinion before risky commandsyes/no guards

  Automation

• AI process auditsplitting a workflow into software, AI and human

  Solo business

• Let Claude work until the goal is metwith /goal

  Automation

• Long-running jobs that wake the model only on changewake gate

  Automation

• Model choicecost per task instead of price per token

  General

• Pre-sorting your inboxa model sorts, Claude drafts, you send

  Automation

• Claude Code in the cloudwhat to clarify first

  Automation

• Collecting web datasettle the access path first, then build

  Automation