deniz.in

Markets

Weather

Loading weather

· via dev.to (home feed)

Anthropic Tells Claude Opus 5.5 Users to Drop 'Think Hard' Prompt Phrases

Anthropic's Opus 5.5 testing found that removing 'think hard' style instructions makes responses start sooner with no clear quality loss, and its new knowledge-work guide covers seven prompting habits to revisit.

Anthropic Tells Claude Opus 5.5 Users to Drop 'Think Hard' Prompt Phrases

Anthropic says thinking prompts are now redundant

On November 12, 2026, Anthropic published a guide to using Claude Opus 5.5 for knowledge work, and a walkthrough of that guide on dev.to pulls out its most practical finding: in Anthropic's own Opus 5.5 testing, deleting instructions such as 'think hard' and 'think carefully' from prompts made replies begin sooner, with no clear drop in quality.

The reason, according to the guide as summarized on dev.to, is that thinking is always on in Opus 5.5. The model adjusts how much it thinks based on the request itself and the selected effort setting, so manual nudges toward deeper reasoning add waiting time without adding reasoning.

Seven habits the guide asks users to revisit

The guide covers seven prompting habits, which the dev.to author worked through using a chart-to-report task.

First, connect the files and apps the task depends on. A prompt for a report built on a spreadsheet and a meeting note should include both, either uploaded directly or reached through a connector to the app where they live. On Team and Enterprise plans, an owner has to enable a connector before members can use it.

Second, define what 'done' looks like in the first message: the goal, the finished output and its audience, and the conditions under which Claude should stop and ask a question. The dev.to author's template covers goal, result, and an ask-me-first-if line for missing data or decisions. In their test, Claude responded with questions about who the chart covered, whether exact figures existed, what an annual data point meant, and who would read the report, which surfaced assumptions before they were baked into the output. The finished file still needs checking against the source data.

Third, drop 'think hard', for the reasons above.

Fourth, start at Medium effort. Anthropic recommends Medium as the opening setting, increasing it only when the model stalls on a task. The control sits behind the model name next to the send button. Before raising effort, the guide advises checking the context, because more thinking time will not supply a missing spreadsheet or an undescribed output.

Fifth, attach the chart instead of typing its numbers. Given a screenshot, Claude described the axes and series, produced a table of approximate values, and noticed that the chart mixed annual, monthly, and quarterly points. The caveat is that values read from an image can look precise in a table while remaining estimates, so exact figures call for the original spreadsheet.

Sixth, name the design choices you do not want. A request like 'Make it look modern' leaves too much open to interpretation; the walkthrough's examples instead specify exclusions, such as no cream background or pill-shaped buttons, alongside requirements like a white background and readable axis labels, with the data unchanged. A reference image also helps when the target layout is already known.

Seventh, share your house style. Attach writing samples, brand guides, or deck templates and say which parts of their style to follow. Feedback should be specific: pointing out that a paragraph repeats the introduction gives Claude a clearer edit than a vague request to make it better.

Specificity beats exhortation

The dev.to post illustrates the thinking-instruction change with two prompts. The weak one stacks thinking instructions around a vague ask, analyzing sales data and reporting what is wrong, without defining the period, the meaning of wrong, or the expected output. The better one names the attached Q3 sales file, asks for the three biggest reasons sales dropped, requests a one-page summary with a simple chart, and tells Claude to stop and ask if data is missing.

The author is careful about what this demonstrates: the screenshots illustrate the prompt change rather than a measured comparison of answer quality, and users should check whether a response actually answers the question and supports its findings with the supplied data.

Why it matters

Habits like 'think step by step' were useful for earlier models and have been carried forward from one release to the next. Anthropic's Opus 5.5 guidance says that era is over for this model: reasoning instructions now cost latency without a quality payoff, and the effort setting is the intended control over how much thinking happens. For developers, the actionable step is auditing existing prompt libraries, removing ritual reasoning phrases and spending that prompt space on deliverables, data sources, and stop conditions instead. One provenance caveat: the details here arrive through a single dev.to walkthrough of Anthropic's guide, so teams should validate the guidance on their own tasks before rewriting production prompts.

  • #prompt-engineering
  • #anthropic
  • #claude
  • #llm

Related posts