SI Agent News
DevelopingClaude

Anthropic advises users to drop old prompting habits for Claude Opus 5.5

Anthropic explains how outdated instructions can slow down its newest model and provides specific tips to improve performance and cost-efficiency.

MediumRecommended starting effort level

The short version

Why it matters

Users can save money and get higher-quality outputs by aligning their communication style with how modern models actually reason. Understanding these shifts helps prevent unnecessary costs when using high-end models like Opus 5.5.

Moving past old habits

Anthropic has released guidance for users of Claude Opus 5.5, noting that many people are still using prompting techniques from older model generations. These outdated habits can actually slow down the model or prevent it from performing at its best. The company explains that because the new model works differently, users need to adjust their approach to get the most out of it.

One specific recommendation is to stop using phrases like 'think hard' or 'think carefully' in prompts. These types of instructions were common a year and a half ago when users had to coax more performance out of early reasoning models. Today, the model is designed to scale its reasoning depth automatically based on the complexity of the task and the selected effort level.

Anthropic's internal testing shows that removing these redundant phrases allows responses to start faster without any measurable drop in quality. The model is now built to understand the requirements of a prompt on its own. By cutting out these old instructions, users can streamline their interactions and see quicker results from the AI.

Optimizing effort and cost

The company suggests that users should start their interactions at the 'Medium' effort level, which is the default for Opus 5.5. In internal tests, the 'Medium' setting on this new model matched the performance of the previous Opus 5 model when it was set to 'High.' This means users can often achieve high-quality results without needing to max out the settings.

Choosing higher effort levels can result in increased costs without necessarily providing a better outcome. Anthropic advises that users should only switch to the 'High' setting if the 'Medium' level fails to meet their needs. This setting can be adjusted manually in the model menu located next to the send button during a chat.

For teams that previously used the cheaper Sonnet model for most tasks and saved Opus for complex projects, Anthropic suggests rethinking that strategy. A more capable model can often complete a task in fewer steps, which helps offset the higher price per token. Users who frequently push Sonnet to 'Extra High' or 'Max' effort may find they are paying nearly as much as they would for Opus 5.5.

Better task management

Anthropic recommends a shift toward describing a desired result in a single message rather than providing a series of one-at-a-time instructions. This approach should include the overall goal, the specific output format, and instructions on when the model should ask follow-up questions. The model is now better at seeing multi-part tasks through to the end without constant human intervention.

The new model is increasingly capable of taking over the back-and-forth communication that users previously handled manually. It can effectively prompt itself from a single initial input to complete a complex workflow. This capability allows the AI to manage the intermediate steps of a project more autonomously than previous versions allowed.

For those who need finished documents or spreadsheets, Anthropic suggests asking for the final product directly. The model often moves from a request to a finished product in a single pass rather than requiring a draft first. This shift in behavior means that users can skip the iterative drafting process for many common professional tasks.

Visuals and design tips

The company highlights significant improvements in how the model reads images, particularly dense charts and tables. Instead of manually transcribing values from a chart into text, users should now attach the diagram as an image or file. Even at the lowest effort level, Opus 5.5 can read these values more accurately than the previous model did at its highest setting.

When it comes to design tasks, Anthropic advises being specific about what to avoid rather than using broad prohibitions. For example, users should name specific styles like 'pill-shaped buttons' or 'cream-colored backgrounds' instead of vaguely banning 'generic design.' General instructions often just result in the model swapping one default style for another without achieving the desired look.

Finally, users who connect Claude to their own files and systems through connectors will see more tailored results. Because the model can pull in the specific information it needs on its own, these connections provide a more accurate context for the AI. This allows the model to ground its answers in the user's specific data more effectively than manual copy-pasting.

What to watch

Try it

Next time you use Claude, try providing your entire project goal, format, and constraints in one single message instead of giving instructions one by one.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by The Decoder · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news