The Demotion Ladder: A Year of Governing Claude Code

Over 14 months the author fed Claude Code a production system of roughly 1,000 API endpoints, generated contracts, Excel/PDF exports, and a shared UI theme. To steer the agent, they accumulated 18 rule files—each 250‑750 lines, totaling more than 5,000 lines of prose. When Claude was asked to build a new CRUD slice, it produced a half‑baked MVP that ignored endpoint conventions, contract hierarchy, generated frontend clients, and UI theming, and even failed to compile. The output resembled work from ten junior developers, highlighting that sheer rule volume does not guarantee compliance. The author cites three preprints: a May 2026 study of 1,650 Claude Code sessions (max 500‑line rule file) that found no effect of rule size on compliance; a June 2026 analysis of 1,323 episodes showing violation rates jump from 0 % to 30‑59 % after the model “compacts” its context; and a 4,416‑trial survey across 12 models revealing that “never do X” prohibitions decay from 73 % compliance at turn 5 to 33 % by turn 16, while “always do X” stays at 100 %. These data confirm three failure modes the author observed: (1) rule relevance is task‑specific, so a vertical slice overloads the agent with competing instructions; (2) repeated context compression strips away detail, requiring costly reminder prompts; and (3) negative instructions erode faster than positive ones, making most of the author’s prohibitions ineffective.

The struggle mirrors a broader industry shift where AI‑assisted coding tools—GitHub Copilot, OpenAI’s Code Interpreter, and emerging “agent” platforms—are moving from free‑form prompt engineering toward structured, deterministic scaffolding. As models grow in capability, token windows remain a hard limit, and developers increasingly embed hooks, test suites, and runtime checks to enforce standards. The author’s “demotion ladder”—prose → louder prose → deterministic hook → pre‑compaction injection—illustrates the practical cost of this transition: each rung adds code overhead (68‑line doc‑selector hook, 24‑line pre‑compaction hook) while still battling the underlying token budget. Competitors that ship built‑in policy enforcement (e.g., Microsoft’s Semantic Kernel with policy layers) may sidestep the need for ad‑hoc hooks, giving them a productivity edge.

Implications include a need to redesign rule delivery mechanisms, favoring concise, positive directives and embedding constraints directly in the generation pipeline rather than relying on large textual manuals. Teams should monitor model release notes for changes in context handling and token limits, and track emerging research on “hook” architectures that survive compaction. Finally, the asymmetry between prohibitions and affirmations suggests that future prompt‑engineering best practices will prioritize “always do X” patterns, potentially reshaping how enterprises codify internal development standards for LLM agents.