The new rules of context engineering for Claude 5 models

Why this is in the vault

First-party guidance from the Claude Code team on how to build harnesses for the Claude 5 generation, published the same day Opus 5 shipped. It is the authoritative outside check on the thin-harness/fat-skills design RDCO's whole agent stack rests on, and it sets the standard the next skill-pruning pass is measured against.

Companion piece to the Claude Opus 5 launch (same day). Headline: Anthropic removed ~80% of Claude Code's system prompt for Opus 5 / Fable 5 with no measurable loss on coding evals. They were over-constraining the model; newer models resolve conflicting guidance ("leave documentation as appropriate" vs "DO NOT add comments") with judgment.

The then→now table (verbatim distillation)

1. Give Claude rules → Let Claude use judgment. Worst-case guardrails ("never write multi-line comments") deleted in favor of "write code that reads like the surrounding code."
2. Give Claude examples → Design interfaces. Examples constrain the exploration space; expressive tool parameters (e.g. a status enum) hint usage better than examples.
3. Put it all upfront → Progressive disclosure. Verification/code-review guidance moved into selectively-loaded skills; deferred-loading tools via ToolSearch. Same applies to CLAUDE.md/SKILL.md: a tree of files loaded at the right time, not a central repository of every practice.
4. Repeat yourself → Simple tool descriptions. Instructions live in tool descriptions, not duplicated in the system prompt.
5. Memory in CLAUDE.md → Auto-memory. Claude saves relevant memories automatically now.
6. Simple specs → Rich references. Plans as HTML artifacts, specs as test suites or reference code, rubrics + verifier agents to encode taste.

Assembly guidance

• System prompt = product context (only matters if you build your own harness).
• CLAUDE.md = lightweight; spend tokens on codebase GOTCHAS, not the obvious; progressive-disclose the rest via skills.
• Skills = lightweight opinion/knowledge carriers particular to you/your team; avoid over-constraint except in highly important areas; split long skills into many files.
• New tooling: claude doctor / /doctor in Claude Code rightsizes skills and CLAUDE.md automatically.

Mapping against Ray Data Co

(filed by Ray, 2026-07-24)

• Validates the thin-harness/fat-skills thesis the shop is built on — from the source, with the 80%-deletion receipt.
• Our CLAUDE.md hard rules survive the test: they encode founder-specific operational gotchas (clock drift, channel-delivery discipline, calendar UTC parsing, subagent routing) — exactly the "particular opinions/knowledge" category Thariq says to KEEP. The pruning target is over-constraining prose in older skills, not the hard rules.
• Queued follow-ups: (1) run /doctor when it lands in our CLI version; (2) rightsizing pass over ~60 skills + CLAUDE.md non-hard-rule sections under these rules — idle-cycle work, pairs with the /improve loop.
• Same-day theme with the Opus 5 daily-driver decision: trust model judgment more; spend scarce constraints (tokens, rules, Fable capacity) only where they're truly yours.

Related

• [[2026-04-15-thariq-claude-code-session-management-1m-context]] — his prior guidance, the source of CLAUDE.md hard rule 4 (route long artifacts through subagents)
• [[2026-07-18-agent-brigade-v2-simplification-design]] — the same "delete coordination, keep domain tools" instinct applied to our own plugin house
• Also: harness-engineering book notes · the /improve skill loop (where the rightsizing pass will run)