The 6 New Rules Router Kit — Anthropic Rewrote Claude Code's Rules; Here's Exactly What to Change Today

The companion kit to the video. On July 24, 2026 Anthropic published 'The new rules of context engineering for Claude 5 generation models' (by Thariq Shihipar, a member of Anthropic's technical staff) and quietly retired the playbook everyone copied for two years: for Claude Opus 5 and Claude Fable 5 they removed over 80% of Claude Code's system prompt — with no measurable loss on their coding evals. INSIDE: (1) ALL 6 SHIFTS AS A CHECKLIST, each with the OLD habit, the NEW rule, and the exact thing to do today — many rules to trusted judgment (they deleted 'default to writing no comments. Never write multi-paragraph docstrings' in favor of 'Write code that reads like the surrounding code'), example dumps to expressive interfaces, front-loaded context to progressive disclosure, repeated instructions to deduped tool descriptions, manual # memory saves to automatic memory, and plain-markdown specs to HTML/code references. (2) THE ROUTER TEMPLATE — the copy-paste thin-router CLAUDE.md skeleton (project one-liner, quick-start commands, max 5 non-negotiables, and a routing table pointing at docs/agent/*.md topic files) plus the 4-step one-evening conversion. (3) OUR MEASURED RECEIPT — a real open-source repo's 46,627-byte CLAUDE.md loaded as 17.7k tokens of memory files every single session (39% of the entire 45.6k startup context); converted to a 1,915-byte router with the content preserved verbatim in 11 topic files, /context reported 845 tokens — a 95% cut, 16.9k tokens saved every session, measured with Claude Code's own /context command on Aug 5, 2026. Plus proof it scales: our production memory system is 545 topic files (~1.92 MB of markdown) with only a ~19 KB router index loaded per session. (4) THE /doctor PAGE — what the agentic health check in Claude Code v2.1.220 actually does (read-only scan first, plain-language report, confirms before changing anything), how to run it, and what's safe to approve. Every shift attributed to Anthropic's post; every number measured on our own machine. Independent kit from Hyperautomation Labs — not affiliated with Anthropic.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.