Anthropic published new prompting guidance for Claude Opus 5.5 saying that old "think carefully" style instructions and effort settings can behave differently now, and may hurt performance instead of helping. This kit gives you the exact before/after rewrites and a checklist to retest your prompts today, so your Opus 5.5 outputs stay consistent.
What changed
The guidance was released alongside the model and covers how to write prompts and set effort levels for it, according to Anthropic's own documentation as reported by Search Engine Journal.
This is a change from prior model generations, where those phrases were commonly recommended to improve reasoning quality.
Effort settings that were tuned for earlier Claude models may not carry over cleanly to Opus 5.5 and could produce inconsistent results if left unchanged.
Search Engine Journal reports that prompts already in production, including chat prompts, may need review under the new guidance, not only prompts written from scratch for Opus 5.5.
Audit your existing prompts first
You cannot fix what you have not listed. Put them in one sheet with columns for prompt text, effort setting, and where it lives.
WhereClaude › Console › Workbench (or wherever your prompts are stored, e.g. a prompts repo or CMS)
These are the exact style of instruction Anthropic's guidance singles out as potentially harmful for Opus 5.5. Flag every match for rewrite.
You will need this baseline to compare before/after outputs once you retest, so you know if a change in output came from the prompt rewrite or the effort setting.
Before/after rewrites you can copy
Replace vague instructions with specific, checkable instructions.
Vague instructions to think carefully do not tell the model what carefully means. Replace with something the model can act on directly.
Before: Think carefully before answering. After: List the [3] factors that matter most for this decision, then give your recommendation in one sentence.
Give the model steps to follow instead of a general request for patience.
Before: Take your time and make sure your answer is correct. After: Check your answer against these [3] criteria: [criterion 1], [criterion 2], [criterion 3]. Revise if any criterion fails, then give the final answer.
Thoroughness is subjective. A required structure forces completeness without relying on the model to interpret 'thorough.'
Before: Be thorough in your analysis. After: Cover these sections in order: [Summary], [Risks], [Recommendation]. Each section must have at least [2] sentences.
Open-ended step by step requests can cause meandering. Defining the steps yourself keeps the model on task.
Before: Think step by step. After: Follow these steps: 1) Identify the goal. 2) List constraints. 3) Propose one solution. 4) State one risk.
Retest and set effort levels
This is where you control how much computation the model applies per response, and it needs to match your rewritten prompts.
WhereClaude › Console › Workbench › Model settings › Effort
Look for the point where output quality stops improving. That is your target setting, going higher just adds cost and latency without benefit.
Test prompt: Summarize the following [document/customer message] in [3] sentences and flag any [risk/complaint]. [Insert content here]
Check for three things: is the answer still correct, is it the same length or shorter, and did it skip the vague filler language. If any of these fails, adjust the rewrite.
Anthropic's guidance may update again. A dated log lets you know quickly which prompts still need retesting next time.
Prompt name: [name] Effort setting: [low/medium/high] Last tested: [date] Rewrite applied: [yes/no] Notes: [any output changes observed]
What good looks like
If you removed a vague phrase and replaced it with a concrete task, checklist, or required structure, that is the correct fix.
Simple lookups or short summaries should use lower effort settings, complex multi-step reasoning tasks should use higher ones. Test both ends of the range to confirm you are not overpaying or underperforming.
This is your proof that you retested rather than assumed the old prompts still work the same way on Opus 5.5.
Want me to look at yours?
If you have gone through this and want a second pair of eyes, book a free call. We open the account together and look at what is actually happening. No deck, no pitch.
A new free sheet most days. Get the next one: follow @ihamza.jameel on Instagram · YouTube
More free sheets
- RETENTIONAI Chat Privacy ChecklistNew · Checklist · 21 checks · 18 Nov 2026
- FILTERSSearch Console New Filters ChecklistNew · Checklist · 14 checks · 16 Nov 2026
- HUBChatGPT Business Hub Setup ChecklistChecklist · 13 checks · 11 Nov 2026
- TRACKINGGoogle Ads Tracking Check: When Conversions DropChecklist · 21 checks · 11 Nov 2026
