Stop Rule-Listing: The Six Context Shifts That Cut Claude Code's Internal Prompt by 80%

TL;DR: Anthropic deleted over 80% of Claude Code's internal system prompt for Opus 5 and Fable 5 without losing quality on coding benchmarks — six shifts worth applying to your own CLAUDE.md and agent definitions.

If your CLAUDE.md has grown into a dense rule list, you may have optimized it for an older model that needed guardrails. Claude 5-generation models behave differently: explicit constraints often conflict with each other and waste tokens the model could use for reasoning. Run /doctor (or claude doctor at the terminal) first — it audits your context setup and flags what to cut.

The six shifts, from Anthropic's July 24 guide:

1. Rules → judgment. Replace "default to writing no comments" with "match the comment density of the surrounding code." Claude 5 infers the intent; spelling out the rule adds noise and creates instruction conflicts.

2. Examples → interface design. Instead of worked examples in the system prompt, design tool parameters that signal their own expected usage. Richer parameter names replace boilerplate examples.

3. Upfront context → progressive disclosure. Move heavyweight skill instructions (code review checklists, verification loops) out of the always-loaded system prompt and into separate skill files that load only when the task needs them.

4. Repetition → single location. Each instruction belongs in exactly one place — the system prompt or a tool description, never both. Duplication registers as a conflicting instruction.

5. Manual memory → auto-memory. Let the model automatically save relevant context rather than maintaining a ## Facts section in CLAUDE.md by hand.

6. Prose specs → rich references. Instead of a markdown summary of your design system, @-reference the actual component files, a test suite, or a rendered screenshot. Models make better inferences from artifacts than from descriptions of them.

In practice: one team applied these to a fleet of specialist agents and shrank definitions from 489 lines each to 37 on average, with measurably better plan-following — per an independent real-world test that also noted where the approach breaks.

What does not work cleanly:

• Auto-memory can pollute across sessions — the model may carry irrelevant context from a prior task into the next one. Experienced developers can catch it; it is a real failure mode.
• "Give Claude judgment" is harder to audit than an explicit rule when something goes wrong. Practitioners on HN reported Opus 5 making harder-to-spot mistakes under the new framing — mistakes that "sound plausible" rather than obviously wrong.
• The 80% reduction was measured on Anthropic's own internal coding benchmarks, not on external codebases; results on a custom repo warrant independent re-evaluation.

Sources: The new rules of context engineering for Claude 5 generation models (Anthropic, July 24 2026), independent real-world analysis with community feedback (Developers Digest)