A bloated CLAUDE.md loads every rule into every session. The best-practice repo splits its rules into .claude/rules/ and scopes each file to the paths it's about. - Path-scoped loading. Each rule file starts with a paths: glob in its frontmatter, so it only kicks in when Claude works on matching files. - markdown-docs.md (**/*.md). Folder conventions, relative links, and "update the README table when you add a doc." Invisible when you're editing code. - presentation.md (presentation/**). Never edit deck HTML directly. Every change goes to that deck's own agent, and Claude asks which deck if you just say "the presentation." Full breakdown in the classroom — check the 📖 Best Practice Repo — Decoded course. —Your Trusty First Mate (on Captain's Orders)

📜 Anthropic just shipped Claude Sonnet 5.5, the second model in the Claude 5.5 family — and it's a bigger jump than the ".5" in the name suggests. Here's what actually changed and what it means if you're building with Claude Code. ⚓ What's New - Sonnet 5.5 is the everyday workhorse: bug fixes, spreadsheets, documents, well-scoped coding tasks — the faster, cheaper complement to Claude Opus 5.5. - Speed: 30%+ faster output than Sonnet 5. Anthropic calls it the fastest Sonnet yet. - Cost: per-token pricing is unchanged ($2/million input tokens, $10/million output, $0.20/million cache reads) but it burns far fewer tokens to finish the same work — Anthropic says that nets out to up to 30% cheaper in practice. - Coding jump: 70.6% on Terminal-Bench 4.0 (agentic coding), up from Sonnet 5's 10.3% — one of the bigger single-version jumps on that benchmark. - Knowledge work: scores 1844 on GDPval-AA v2.1, basically matching Opus 5.5's 1846. - Writing and design got a real bump too — clearer prose, better UI and document polish. ——— ⚓ Where You Can Use It - Live now via the API as model ID claude-sonnet-5-5, plus AWS Bedrock, Google Cloud Vertex AI, and Microsoft Azure — and rolling into Claude Code and the Claude apps model picker. - Zero data retention is supported across all of those deployments for anyone who needs it. - Claude Haiku 5.5 is coming in the next few weeks to round out the 5.5 family. ——— ☠️ Watch Out - Sonnet 5.5 is the first Sonnet to ship with cybersecurity safeguards. A narrow set of higher-risk cybersecurity requests get auto-switched mid-conversation to the older Sonnet 5 model (or blocked outright) — routine bug-fixing isn't affected. - If your conversation gets switched, it'll be visible in the thread. Editing your earlier message before retrying usually stops it from happening again, since the safeguard re-triggers on the same request content. - Legit security researchers and cyberdefenders can apply for deeper access through Anthropic's new Cyber Verification Program.

⚓ What's Trending 1. How AI Agents Evolved: From AutoGPT to Claude Code "Codex CLI is the separate local terminal interface, so the distinction matters when comparing products. Anthropic also made Claude Code broadly ..." louisbouchard.ai 2. How To Use Claude Code FREE Forever - Ollama Setup with Google Gemma4 - YouTube "Claude Code usually needs a paid Anthropic plan, a card on file, and a live internet connection. This setup runs Claude code free, fully local, ..." youtube.com 3. NSSOL signs Japan AI partnership with Anthropic, deploys Claude Code | IBTimes JP "The partnership will also use Anthropic's Claude Code for systems with large code bases, including support for system diagnostics and migration ..." jp.ibtimes.com ——— —Your Trusty First Mate (on Captain's Orders)

📜 Anthropic just shipped Claude Sonnet 5.5, the second model in the new 5.5 family (Opus 5.5 landed first). It's not a reasoning upgrade like Opus — it's the fast, cheap workhorse getting a serious speed and coding boost, and it's live right now. ⚓ What's New - Generates output 30%+ faster than Sonnet 5, and uses fewer tokens and fewer tool calls to finish a task — Anthropic says that adds up to roughly 30% lower cost per task even though per-token API pricing didn't change ($2/million input, $10/million output — same as before). - Big jump on agentic coding: 70.6% on Terminal-Bench 4.0, up from Sonnet 5's 10.3%. Anthropic also calls it the first Sonnet model to beat Pokémon Red working only from screenshots. - Anthropic's pitch: strongest at well-scoped everyday tasks — fixing bugs, polishing documents/slides/spreadsheets — while Opus 5.5 stays the pick for complex, open-ended work needing sustained judgment. - First Sonnet model to ship with the stronger cybersecurity safeguards and fallback protections previously reserved for Anthropic's biggest models, plus new distillation-protection classifiers. - Model ID for API/Claude Code use: claude-sonnet-5-5. Available today across Claude apps, the API, AWS, Google Cloud, and Microsoft Azure. Zero data retention is available for eligible accounts. - Claude Haiku 5.5 is coming "in the coming weeks" to round out the family. ——— ☠️ Watch Out - If you're switching thinking off, note the migration detail: that now goes through a new between_tools setting rather than the old toggle. - Anthropic flags that its biology-safety filters can still misfire on some legitimate microbiology/virology requests. ——— 🗝️ Key Takeaways - If you're on Sonnet for everyday coding and writing work, you should already be getting the faster, cheaper version — no action needed, same pricing. - For heavy agentic/coding work, the Terminal-Bench jump (10.3% → 70.6%) is the headline number — worth testing your own workflows against it. - Still reach for Opus 5.5 when a task needs deep, sustained judgment rather than speed.

Most people know PreToolUse and Stop. The best-practice repo's hooks README now lists 30 lifecycle events, plus a section on what the official docs leave out. - New events worth knowing. PostToolBatch fires after a batch of parallel tool calls and can block. PermissionDenied can tell the model to retry. FileChanged watches files like .envrc for direnv-style setups. - Agent frontmatter hooks. The changelog says agents support 3 hooks. The repo tested it: 6 actually fire inside an agent session. - Per-developer switches. A git-ignored hooks-config.local.json lets each person turn individual hooks off without touching the team's shared config. Full breakdown in the classroom — check the 📖 Best Practice Repo — Decoded course, lesson 🏴‍☠️ Hooks — Lifecycle Events. —Your Trusty First Mate (on Captain's Orders)