Anthropic Says Your Prompting Habits Are Slowing Down Claude Opus 5.5
The prompt engineering incantations that made earlier chatbots work are now dead weight, according to Anthropic. In guidance published Nov 12, the company told users of Claude Opus 5.5, its flagship model that launched Sept 22, 2026 and now serves as the default on Pro, Max, Ultra, and Team plans, that many of them are holding the model back with habits carried over from older generations.
The most pointed advice: stop typing "think hard" or "think carefully." Opus 5.5 reasons on its own for every request and scales its reasoning depth to the task and the selected effort level, so the magic words add nothing. In Anthropic's tests, responses without those phrases started faster and showed no measurable drop in quality. The practice dates back roughly a year and a half, to early reasoning models whose internal routers users could sometimes nudge with the right phrasing. That machinery no longer responds to the tricks.
The effort knob is the new prompt
Instead of verb spells, Anthropic wants users to work the effort setting. Start at Medium, the default. The company says Opus 5.5 at Medium matched Opus 5 at High in internal tests, and that cranking higher settings costs more without reliably improving results. Switch to High only when Medium falls short. The control sits in the model menu next to the send button.
The economics matter here. Anthropic's cost documentation notes that teams running cheap Sonnet models for routine work while reserving Opus for hard tasks should rethink the split. A more capable model often finishes in fewer steps, which partly offsets its higher per-token price. Sonnet pushed to Extra High or Max effort ends up costing about as much as Opus 5.5 anyway. Paying for effort on a weaker model buys the price without the capability.
Describe the destination, not the route
The broader shift Anthropic describes is from step-by-step chatting to goal specification. Rather than drip-feeding instructions one message at a time, users should state the desired end result in a single message: the goal, the output format, and when the model should pause and ask questions. Opus 5.5 sees multi-part tasks through, Anthropic says, and effectively prompts itself from that initial input, sometimes handling the back-and-forth better than the user would have.

Several related habits drop away under that model. Attach diagrams and tables as images or files instead of transcribing their values into text; the workaround existed because older models read charts poorly. Anthropic says Opus 5.5 at its lowest effort level read values from dense charts more accurately than Opus 5 at its highest. Ask for the finished document or spreadsheet directly rather than requesting a draft first, because the model often goes from request to finished product in one pass. And connect Claude to your own files and systems through connectors, so it can pull the context it needs itself instead of receiving it secondhand.
For design work, the guidance gets oddly specific: name the styles you do not want. Banning "generic design" in the abstract just swaps one default look for another, Anthropic says. Prohibiting cream-colored backgrounds or pill-shaped buttons by name actually moves the output.
Read the advice with the right grain of salt
One caveat deserves emphasis. The claims about Medium matching High, about faster responses without "think hard," and about chart reading come from Anthropic's internal tests. The company published no public benchmark alongside the guide. The recommendations are still sensible starting defaults, and the cost logic around Sonnet at max effort traces to Anthropic's own pricing documents, but independent eval numbers are absent.
The larger story is what the guidance admits about how fast this layer of skill moves. Prompt engineering as a discipline is barely two years old, and Anthropic is now telling its own users that a meaningful chunk of it has expired, that the tricks which coaxed performance out of 2025-era reasoning models have become noise at best and drag at worst. As models absorb more of the scaffolding users used to bolt on, the remaining skill shifts toward stating goals, constraints, and formats clearly, and knowing which knob to turn. The craft is becoming management rather than spellcraft.
