Opus 5 で壊れるプロンプト3つ — 公式が「消せ」と言った指示たちThis article identifies three prompt instruction patterns that Anthropic…

As development teams begin planning migrations to Claude Opus 5, Anthropic's model-specific guidance has drawn attention for an unusual reason: it recommends deleting several prompt instructions that were once treated as best practice. A recent Zenn write-up distills that guidance into three instruction patterns that are likely to behave differently, or stop working as intended, once existing prompts are pointed at the newer model. The risk matters because it is subtle. Prompts that contain these patterns do not throw errors; they simply drift away from expected behavior, which can be harder to catch than an outright failure.

The first pattern involves explicit chain-of-thought scaffolding — instructions such as "think step by step" or lengthy directives telling the model exactly how to reason before it answers. Opus 5 appears to handle much of this reasoning internally, so manually forcing an externalized thought process can now be redundant or even counterproductive. According to the guidance, hand-written reasoning instructions may produce padded output or interfere with the model's own planning. Where visible reasoning is genuinely required, the recommendation is to rely on the model's native reasoning controls rather than improvised prompt text.

The second pattern concerns heavy-handed emphasis: capitalized commands like "YOU MUST," instructions repeated several times for reinforcement, and urgency or threat framing meant to force compliance. Earlier models sometimes needed this kind of pressure to follow constraints reliably. With stronger instruction-following in Opus 5, such emphasis is described as unnecessary, and in some cases it can skew the model toward overly cautious or excessively literal responses. The suggested fix is straightforward: state each instruction once, clearly, and trust the model to follow it.

The third pattern targets overly rigid output and role constraints — verbose persona preambles, redundant formatting rules, and negative instructions such as "do not hallucinate" or exhaustive lists of prohibitions. These can crowd the context window, dilute the instructions that actually matter, and, in the worst case, cause the model to over-index on avoidance rather than on completing the task. The guidance frames leaner system prompts as more robust, since the model is less likely to latch onto a stray constraint at the expense of the primary goal.

Taken together, the three patterns reflect a recurring theme in Anthropic's documentation: as models improve, prompts written to compensate for older weaknesses can become liabilities. This is not unique to Claude. OpenAI's reasoning-oriented models have carried similar advice, noting that explicit "let's think step by step" prompts add little value when reasoning is built into the model, and Google has offered comparable notes for its Gemini line. The broader industry shift is toward prompts that specify intent and constraints while leaving the mechanics of reasoning to the model.

For readers less familiar with the terminology, a few prerequisites help. A system prompt is the persistent instruction block that frames how a model behaves across a conversation. Chain-of-thought refers to prompting a model to produce intermediate reasoning steps, a technique that measurably improved older models but is increasingly native to newer ones. Instruction-following describes how reliably a model adheres to stated constraints, which tends to improve with each generation and reduces the need for emphatic phrasing.

The practical takeaway from the article is that migration should be treated as a testing exercise rather than a drop-in swap. Because these patterns fail quietly, the recommended approach is to build a small evaluation set that captures the behaviors a workflow depends on, run existing prompts against Opus 5, and compare outputs before and after trimming the deprecated instructions. Regression testing of this kind is becoming standard practice as teams move between model versions, and tooling around prompt evaluation has grown accordingly.

It is worth noting that the specifics here are drawn from a third-party summary of Anthropic's guidance, so developers should confirm the details against the official migration notes for their exact use case. The underlying principle, however, is durable: prompts accumulate workarounds over time, and a model upgrade is a natural moment to remove instructions that a more capable system no longer needs.