Claude Code Just Changed — the 6 New Rules of Context Engineering

TL;DR

Anthropic engineer Tariq published "The new rules of context engineering for Claude 5 models" (4M+ views), and several of the rules are the opposite of the advice we followed for months — because the models got dramatically smarter (Anthropic removed 80%+ of Claude Code's system prompt for Opus 5 / Fable 5 with no eval loss; the artificial-analysis intelligence score jumped from ~31% a year ago to ~60% now). The shift: stop over-constraining the model and instead organize your context well and let the model use judgment. Jay frames context with his ARMS model (Applications, Routines, Memory, Skills) and shows how to adopt all six rules — thin routing CLAUDE.md, design systems instead of examples, HTML references, and self-updating skills like /calibrate, plus a free /doctor-plus to audit your setup against the new rules.

The 6 rules (then → now)

#	Before	Now
1	Give Claude lots of rules	Let Claude use judgment
2	Give it examples	Give it design interfaces (a system)
3	Put all context up front	Progressive disclosure (load context when needed)
4	Repeat yourself	Simple tool descriptions (no duplication)
5	Manually save memory (#)	Automatic memory
6	Simple specs (markdown)	Richer references (HTML artifacts)

The framing: context = ARMS

Your prompt is only a sliver of what Claude works from; the rest is your context, which Jay organizes as ARMS:

flowchart LR
    A["Applications<br/>tools via MCP / API / CLI"] --> CTX(("Your context"))
    R["Routines<br/>crons / scheduled tasks"] --> CTX
    M["Memory<br/>artifacts, reports, docs"] --> CTX
    S["Skills<br/>SOPs invoked by /command"] --> CTX
    CTX --> OUT["Engineer these well →<br/>ahead of 99% of agent users"]

---

Rule 1 · Let Claude use judgment

Old system prompts were stringent ("write no comments, never multi-paragraph docstrings"). The new one trims to essentially "write code that reads like the surrounding code — match its comment density, naming, and idiom." For creative work, let Claude run multiple options instead of caging it. Jay built a surprise-me skill that tells Claude to "demonstrate extreme capability, taste, and artistic flavor" — producing front-end designs that dodge the AI-slop look.

Rule 2 · Design interfaces, not examples

Examples constrain the model's exploration space. Instead, give it a design system. Jay's /robo skill points to a brandbook.html (color palette, voices, fonts, his dot-matrix style) so every asset is on-brand with freedom inside the guidelines.

Adopt it: ask Claude for several design systems → give feedback until you love one → have it write a brandbook.html of the rules → turn that into a skill (/robo, /helvetica). Every future asset looks that good.

Rule 3 · Progressive disclosure — CLAUDE.md as a router

Stop stuffing CLAUDE.md with every practice. Make it a thin router to a tree of files that loads context only when a session needs it. Jay's "second brain" (57,000 files) uses CLAUDE.md to route by department (content, community, product, personal, business), with sub-routers (e.g. content.md) pointing to the right skills/files per intent.

flowchart TD
    CM["CLAUDE.md — thin router"] --> C1["content.md → content skills/files"]
    CM --> C2["community.md → community files"]
    CM --> C3["product.md → dev files"]
    CM --> C4["business.md → ops files"]
    C1 --> S1["only loads when you work on content"]

Why it matters: a thick CLAUDE.md burns those tokens every session from the first prompt; a thin router saves tokens (and usage-limit headroom) that compound the more you use Claude.

Rules 4–5 · Simpler tools + automatic memory

• Rule 4 — simpler tool descriptions: smarter models no longer need repeated instructions in both the system prompt and the tool description; kill the duplication and save tokens.
• Rule 5 — automatic memory: Claude Code now auto-saves relevant memories. Jay still captures the good stuff deliberately with a /calibrate skill — run at session end, it reviews the conversation and updates skills, CLAUDE.md rules, memory, and workflows, so the system learns you over time.

Rule 6 · Richer references (HTML > markdown)

Markdown is still core (CLAUDE.md, skills), but newer models handle richer references — and HTML artifacts win when the content is visual (a brandbook's color palette can't live in markdown). Bonus: you can read HTML infographics far faster than a wall of terminal text. Jay routinely has Claude turn concepts into HTML infographics via his design skill.

Adopt it fast: /doctor and /doctor-plus

Latest Claude Code ships /doctor — checks your install (broken/duplicate installs, path problems), finds dead weight, trims a bloated CLAUDE.md, flags slow hooks, and reports before applying. Jay found it a bit basic, so he made a free /doctor-plus that also audits your skills/artifacts against the six shifts (e.g. flagging a 2,090-line skill that should be a router) — good as a monthly routine.

---

My adoption checklist (for the Hermes/CKA setup)

1. Thin the CLAUDE.md → make it a router; move detail into content.md / department sub-routers.
2. Build a brandbook.html design interface + a /brand skill (CKA voice, colors) so assets are on-brand with freedom.
3. Add a /calibrate skill to capture each session's learnings into skills + memory automatically.
4. Prefer HTML artifacts for anything visual or worth reviewing at a glance.
5. Let the model judge on creative work — ask for multiple options instead of over-specifying.
6. Run a monthly /doctor-plus to keep the setup lean and token-efficient.