[{"data":1,"prerenderedAt":45},["ShallowReactive",2],{"$fCsg8FZwcwXulHB3aXPEz3ifUmFp-Nq9d7F8Fqq1H48g":3},{"date":4,"generated_at":5,"picks":6,"candidates_scanned":44,"candidates_scored":42},"2026-04-19","2026-04-19T06:00:00.000000+00:00",[7,21,35],{"rank":8,"title":9,"source":10,"url":11,"category":12,"tldr":13,"score":14,"scores":15,"why":20},1,"Changes in the system prompt between Claude Opus 4.6 and 4.7","Simon Willison","https://simonwillison.net/2026/Apr/18/opus-system-prompt/#atom-everything","Guide","- Simon Willison diffed the official Claude.ai system prompts between Opus 4.6 and 4.7 — Anthropic publishes these, but the changes aren't highlighted in the launch blog post\n- Newly added: \"Claude in PowerPoint — a slides agent\" is now listed as a tool Claude Cowork can use, alongside Claude in Chrome and Claude in Excel (not mentioned in 4.6)\n- The child safety section was greatly expanded and now lives inside a `\u003Ccritical_child_safety_instructions>` tag — once Claude refuses on child safety grounds, it must treat all subsequent messages in the same conversation with \"extreme caution\"\n- A new `\u003Cacting_vs_clarifying>` section explicitly tells Claude: \"the person typically wants Claude to make a reasonable attempt now, not to be interviewed first\" — Claude should only ask clarifying questions when the request is literally unanswerable without the missing info\n- Claude is now instructed not to keep users engaged when they signal they want to stop — \"does not request that the user stay in the interaction or try to elicit another turn\"",67,{"direct_claude_relevance":16,"practical_utility":17,"novelty":18,"source_credibility":19},28,14,13,12,"Simon Willison used Claude Code to reconstruct Anthropic's published system prompt history as a proper git timeline, then diffed Opus 4.6 → 4.7. The result surfaces several behavioral changes that weren't in Anthropic's launch announcement: a new slides agent (Claude in PowerPoint), an explicit instruction to act before asking questions, and a substantially expanded child safety enforcement model. Anyone building Claude integrations or trying to understand why 4.7 behaves differently from 4.6 will find this more useful than the launch post itself.",{"rank":22,"title":23,"source":24,"url":25,"category":26,"tldr":27,"score":28,"scores":29,"why":34},2,"Tired of Claude Code Ignoring Your CLAUDE.md? I Built claudemd-guard","Dev.to Claude","https://dev.to/hirochiba/tired-of-claude-code-ignoring-your-claudemd-i-built-claudemd-guard-34k5","Tutorial","- In long Claude Code sessions, CLAUDE.md rules get pushed out of the context window by context compression — the rules become invisible to the model and it starts violating them\n- claudemd-guard is a PreToolUse hook that intercepts every Edit, Write, and Bash call before it runs, checks the operation against your CLAUDE.md rules using Claude itself, and blocks violations before they happen\n- Install with two commands: `npm install -g claudemd-guard && claudemd-guard install` — then restart Claude Code. No extra API key needed (runs via your local Claude CLI). Fail-open design: if the guard itself errors, operations are allowed through",53,{"direct_claude_relevance":30,"practical_utility":31,"novelty":32,"source_credibility":33},18,20,10,5,"Context compression silently killing CLAUDE.md enforcement is one of the most common Claude Code frustrations for projects with many strict rules — and there's been no built-in solution. claudemd-guard attacks the problem at the tool-call level rather than hoping the model remembers, which is the right place to enforce it. The fail-open design is a sensible choice for a convenience tool: stopping all work because the guard errored is worse than letting a borderline operation through. Two-command install, no API key, uninstall with one command.",{"rank":36,"title":37,"source":24,"url":38,"category":12,"tldr":39,"score":40,"scores":41,"why":43},3,"dotclaude: The Open-Source Governance Layer for AI-Assisted Development","https://dev.to/kaiohenricunha/dotclaude-the-open-source-governance-layer-for-ai-assisted-development-3177","- dotclaude is an MIT-licensed monorepo with two uses: (1) personal — symlinks a portable skills + slash-commands library into `~/.claude/` so every Claude Code session on your laptop has the same tools, and (2) team — a governance CLI that enforces spec-backed PRs, skill-manifest integrity, and architectural drift detection in CI\n- The personal bootstrap is three commands: `git clone https://github.com/kaiohenricunha/dotclaude.git ~/projects/dotclaude && cd ~/projects/dotclaude && ./bootstrap.sh` — after that, every repo automatically gets cloud/IaC specialist skills (AWS, GCP, Azure, Kubernetes, Terraform, Pulumi) that auto-trigger when you mention the relevant technology",52,{"direct_claude_relevance":31,"practical_utility":42,"novelty":32,"source_credibility":33},17,"As teams adopt Claude Code at scale, the lack of shared discipline becomes a real problem: five engineers each with their own prompt tricks, no consistent review depth, no enforcement of spec alignment. dotclaude solves both the individual portability problem (skills travel with you across repos) and the team consistency problem (CI-enforced PR governance). The dual-persona monorepo design means you can start with the individual path and grow into team governance without switching tools.",25,1776575098212]