Introduction

If you still add the phrase think hard to the end of every Claude prompt, you may be carrying an old habit into a new model. Anthropic released guidance for Claude Opus 5.5 on November 12, 2026, and it lists seven prompting habits worth revisiting for knowledge work. Developer and AI researcher Vladislav Guzey walked through the recommendations, testing them with real chart-to-report tasks. This post breaks down what changed, why it matters, and how to apply it even if you have not seen the original walkthrough.

What Happened

Anthropic's November 2026 guide covers seven areas where prompting behavior should shift for Opus 5.5. The core idea is that the model handles reasoning internally, so users should focus on context, clarity, and output definition rather than urging the model to think harder.

  • Connect the files and apps the task depends on before asking for a result. On Team and Enterprise plans, an owner must enable connectors before team members can use them.
  • Describe what done looks like in the first message: the goal, the finished format and audience, and when Claude should stop to ask a question.
  • Drop instructions such as think hard or think carefully. Thinking is always enabled for Opus 5.5, and the model adjusts depth based on the effort setting.
  • Start at Medium effort and increase it only if the model stalls. Extra thinking cannot replace a missing spreadsheet or an undefined output.
  • Attach charts, screenshots, or spreadsheets instead of typing numbers by hand. Values read from an image are estimates and should be verified.
  • Name specific visual styles to avoid instead of using vague terms like modern.
  • Share samples of your house style and give sentence-level feedback rather than general directions such as make it better.

In one example, Guzey asked Claude to turn a chart into a one-page management report. Claude paused to ask who the chart covered, whether exact figures were available, what an annual data point represented, and who would read the report. After receiving those answers, it produced a summary and a chart. The author notes the screenshots illustrate the prompt change and are not a measured comparison of answer quality.

Why This Matters

Prompting habits formed with earlier models do not always transfer cleanly. Telling Opus 5.5 to think carefully can actually delay the start of a response without improving quality, according to Anthropic's testing. Vague requests also force the model to guess about data, time periods, and desired formats, which increases the risk of errors in the final output.

Defining sources, output expectations, and stopping points up front reduces back-and-forth. It also makes it easier to check Claude's work against the original files, which remains essential because the model can misread chart labels or mix up monthly, quarterly, and annual data points.

Key Takeaways

  • Upload or connect source files before prompting so Claude is not guessing about missing context.
  • State the goal, the finished file format, the intended audience, and when to ask for clarification in the first message.
  • Remove think hard and think carefully from prompts; use the Medium effort setting as a starting point.
  • Attach charts and spreadsheets directly, then verify any numbers Claude extracts from images.
  • Give concrete design constraints, such as no cream background, instead of abstract style directions.
  • Share writing samples or brand guides and point to specific sentences or slides that need revision.

Conclusion

Pick a real report, document review, or chart you already need to complete. Supply the files, describe what finished looks like, start at Medium effort, and check the output against your brief before adjusting the prompt or effort level. The full set of prompts and a video walkthrough are available from the author for readers who want to follow along step by step.