Knowledge map
Each dot is a page, each line a connection. Click a dot to open the page.
- Topic
- News
- Knowledge card
- Quick read
All pages as a list (74)
Topic
News
- Agent Skills: Anthropic hands the skill format to the Linux Foundation
- AI Act for medical devices delayed to August 2028: more time, same work
- AI as a document checker: run a counter-test and a repeat test before relying on it
- AI in regulatory work: grounded in sources, not a text generator
- Claude Design moves into Claude: claude.ai/design closes on 14 December
- Claude in October: Sonnet 4.5 retires on 30 November, Haiku 5.5 is new
- Cloudflare AI Gateway: Auto Router picks the AI model for each request
- Cloudflare Workers AI: AI responses contained other customers' data
- DIGICAP: medical device dossiers are meant to become machine-readable
- FDA agenda for 2027: final AI lifecycle guidance planned, but no date set
- FDA: radiology AI still needs a 510(k), including follow-on devices
- Generative AI medical devices: what FDA has put up for discussion
- GPT-6.1 Terra: OpenAI model at $1 input and $5 output per million tokens
- n8n 2.44 is stable: prompt caching for Claude has to be switched on per node
- New standards for AI: EN 18286 and ISO/TS 24971-2 are out, presumption of conformity is not
- UK government confirms indefinite recognition of CE-marked medical devices
- UK: AI medical devices to be monitored across their whole lifetime
Knowledge card
- "Done" only with evidence – a check command instead of self-reporting
- A second opinion before risky commands – yes/no guards
- AI process audit – splitting a workflow into software, AI and human
- Checks the agent cannot bypass – checkpoints and gates
- Claude Code in the cloud – what to clarify first
- Collecting web data – settle the access path first, then build
- Decision models – when a model should only choose
- Does a skill still help? – measuring skills
- Have your own website checked from the outside
- Keeping context clean – fewer tokens, clearer answers
- Knowledge base for agents – the LLM wiki after Karpathy
- Let Claude work until the goal is met – with /goal
- Long runs and parallel work – Claude Code's built-in tools
- Long-running jobs that wake the model only on change – wake gate
- Maintaining your Claude Code setup – clear out after every model change
- Memory in three tiers – what Claude knows about you and your projects
- Model choice – cost per task instead of price per token
- Pre-sorting your inbox – a model sorts, Claude drafts, you send
- Project map instead of rereading – Claude Code with a code map
- Project start with rules and plan files
- Self-improving loop with locked checks
- Separate building and checking – a critic with fresh context
- Vet third-party skills before they run
- Websites without the AI look – directions, references, checks
Quick read
- "Done" is a command with an expected output
- /goal: a second model decides when to stop
- A long-running job needs no model while nothing is happening
- A map of the code only pays off for large projects
- A model that only decides costs almost nothing
- A setup ages with every new model
- A skill has to prove itself against “without”
- A wiki that the model maintains itself
- An agent can forget a rule, but not a gate
- Cheaper per task is not the same as cheaper per token
- Claude Code in the cloud: privacy first
- Every statement gets a label: seen, suspected, unknown
- Files first, then code
- Interview first, then the draft
- Many agents in parallel – only if the task splits up
- Memory belongs in files, not in the chat
- References instead of adjectives
- Rewinding is cheaper than patching
- The check rules do not belong to the agent
- The decision model sorts, Claude drafts, you send
- Third-party skills are third-party code
- Watch first, then automate
- Whether a scraper gets through also depends on where it comes from (one measurement)
- Whoever builds does not check
- Whoever executes should not sign off their own work