Detailed Analysis
A Reddit post analyzing Anthropic's Claude Code architecture has surfaced an interesting internal tension within the company's own technical documentation, prompting questions about the design philosophy behind Skills, the feature that lets Claude dynamically load specialized instructions and capabilities into its context window. The author, building a product called Writ, juxtaposes two of Anthropic's engineering blog posts: one on Agent Skills, which describes progressive disclosure of skill names and descriptions into the system prompt before Claude decides whether to load the full SKILL.md file, and another on context engineering, which explicitly warns about "context rot" — the degradation of a model's recall accuracy as more tokens accumulate. The apparent contradiction is that Anthropic simultaneously advocates for lean, high-signal context while building a flagship feature that relies on the model itself to judge relevance and pull in additional instructions, rather than enforcing behavior deterministically.
The crux of the critique centers on the distinction between probabilistic guidance and deterministic enforcement. Claude Code already ships with PreToolUse hooks, which can block a tool action outright before it executes, even in bypass-permissions modes. This is a mechanism that operates outside the model's discretion entirely. The author argues that Anthropic clearly understands this distinction, since their own Skills documentation acknowledges that certain tasks require "the deterministic reliability that only code can provide." Yet workflows described as "mandatory" within the Skills system remain, in practice, contingent on Claude noticing a natural-language instruction, loading it, interpreting it correctly, and choosing to comply. That is a fundamentally different guarantee than a hook that physically prevents an action from occurring. The post is careful to frame this as a design-boundary question rather than an accusation, but it does raise the uncomfortable observation that input tokens are billable, and a system that leans on natural-language instruction-following over code-level enforcement will, structurally, consume more tokens than one that doesn't.
This matters because it touches on a broader credibility question facing Anthropic and other AI labs: whether safety- and reliability-oriented technical writing is fully aligned with product incentives. Anthropic has built much of its public identity around rigorous, safety-conscious engineering communication — publishing detailed guides on agentic best practices, context management, and responsible tool use. When that same body of documentation appears to expose an internal inconsistency between stated best practices (keep context minimal, use deterministic code where possible) and actual product defaults (encourage model-driven context expansion), it invites scrutiny of whether business incentives around token consumption are influencing architectural choices, even unintentionally. The author explicitly disclaims having proof of deliberate design intent, which is an important caveat, but the underlying observation, that the people who wrote the context-rot warnings are the same people who shipped a context-expanding feature, is a legitimate point of tension worth examining rather than dismissing.
More broadly, this discussion reflects a maturing skepticism within the developer community about agentic AI tooling as it moves from novelty to production infrastructure. Early enthusiasm about agents that can dynamically reason about their own instructions is giving way to more rigorous engineering scrutiny, with developers increasingly distinguishing between "soft" governance (prompts, instructions, and model judgment) and "hard" governance (code-level enforcement, permission systems, hooks). This mirrors a pattern seen across the industry as agentic systems get deployed in higher-stakes contexts: teams are learning that anything genuinely mandatory, whether a security policy, a compliance rule, or a coding standard, should live outside the model's probabilistic decision-making loop rather than inside it. The fact that a builder like the Writ author is explicitly redesigning their own product to move "objective, important and genuinely mandatory" rules outside the model's context entirely signals a broader shift toward treating LLM agents as components within deterministic systems, rather than as autonomous decision-makers that can be fully trusted to self-govern via instructions alone. Whether or not Anthropic's Skills design was influenced by token-revenue incentives, the episode underscores how the industry is still working out where the boundary between model judgment and hard-coded enforcement should sit, and how much that boundary matters for trust in agentic systems.
Read original article →