# Capable Models Need Fewer Constraints Not More ## Claim As a model's judgment improves, the highest-leverage context-engineering move shifts from adding explicit rules to removing them and designing an environment the model can reason inside. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds for models at or above roughly Claude-5 capability, where internal evaluation reportedly showed no measurable regression after removing ~80% of a system prompt. - Breaks for older or narrow-domain models that lack general judgment — they still need explicit guardrails. - Breaks for security-critical instructions (never commit secrets, sanitize inputs) that should stay explicit regardless of model capability; capability doesn't remove the cost of a rare catastrophic failure. - Does not establish that judgment-based prompting works as well for non-coding domains (writing, analysis, research) where correctness is harder to verify than in code. ## Evidence An Anthropic engineer reported that Claude Code's system prompt was cut by roughly 80% for Claude Opus 5 / Fable 5 with no measurable loss on coding evaluations, after internal transcripts showed accumulated rules actively conflicting with each other ("leave documentation as appropriate" vs. "DO NOT add comments" clashing in the same request). The claim rests on a single vendor's internal benchmark, not an independently replicated study — treat the specific percentage as directional, not verified. ## Links - [[3. Resources/Permanent Notes/Progressive Disclosure Keeps A Plugin's Core Trustworthy]] — because extends: the same progressive-disclosure principle applied to a different surface, moving detail out of the always-loaded path and into what's fetched on demand - [[3. Resources/Permanent Notes/Raw Provenance and Distilled Guidance Serve Different Memory Jobs]] — because applies-in: the practitioner-side version of the same shift, distilling durable guidance out of raw interaction history rather than repeating it every session - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — because applies-in: how capability growth changes the design of the tools and instructions surrounding a model ## Tests / falsifiers - Find a case where trimming an established rule from a Claude-5-class CLAUDE.md or skill measurably degrades output quality on a task the rule was written for — that would bound the claim's scope more precisely than "capable models" alone. - Compare rule-heavy vs. judgment-based versions of the same skill on identical tasks and measure pass rate, not just anecdote. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Literature Notes/The New Rules of Context Engineering for Claude 5 Models]] — reports the 80% system-prompt reduction and the five resulting shifts in context-engineering practice.