Updated 15 hours ago
Anthropic Says Your Claude Prompts Are Outdated — Here's What to Change

Claude Prompt Engineering

Anthropic Says Your Claude Prompts Are Outdated — Here's What to Change

Anthropic published new prompting guidance for Claude Opus 5.5 telling developers to drop 'think hard' and 'think step by step' from their system prompts. The model's always‑on adaptive reasoning makes these legacy instructions redundant — and in some cases actively harmful to speed and cost.

The Big Change: Always‑On Thinking Makes Old Prompts Obsolete

Anthropic dropped a new prompting guide for Claude Opus 5.5 and the headline is simple: your old prompts are slowing the model down. Opus 5.5 uses always‑on adaptive reasoning — thinking cannot be turned off. Setting thinking: {"type": "disabled"} now returns a 400 error, according to Anthropic's platform docs.

The model self‑determines how much to think per request based on task complexity and the new effort parameter. Previous Claude generations let developers turn thinking off and used prose instructions like "think carefully" or "think step by step" to coax deeper reasoning when needed. That era is over.

Anthropic's own testing showed that removing "think carefully" from chat system prompts made replies start faster with no clear decline in reply quality, THE DECODER reported. The company that previously told developers to add "This task involves multistep reasoning. Think carefully before responding" is now telling everyone to strip it out.

What to Drop: The Old‑Habit‑to‑New‑Practice Map

Anthropic's guidance, summarized across the official docs and community analysis, provides a clear before‑and‑after map for updating your prompts:

  • Drop "think hard" / "think carefully" Delete from all system prompts and CLAUDE.md files. Use the effort parameter instead. Per‑message thinking nudges still work for targeted turns — just don't make them standing instructions, ClaudeFA.ST notes.
  • Drop reasoning‑before‑answer requests Drop asking the model to write out its reasoning before answering. This can trigger reasoning_extraction refusals. Read reasoning from thinking blocks instead (Ctrl+O or display: 'summarized'), per ClaudeFA.ST.
  • Drop multi‑turn step‑by‑step chat Switch to single‑message goal‑setting. State the goal, output format, and interruption conditions once. The model self‑prompts from a single input, per.2
  • Drop two‑pass draft‑and‑final requests Drop two‑pass requests that ask for a draft then a final version. Ask for the finished document or spreadsheet directly. The model often goes request‑to‑finished in one pass, UX Planet reports.
  • Drop vague style prohibitions Drop vague style prohibitions. Instead of saying avoid generic AI design, name specific unwanted elements such as cream backgrounds, pill‑shaped buttons, and centered hero text, ClaudeFA.ST recommends.
  • Drop vague style bans The approach is to name specific unwanted design elements rather than issuing broad prohibitions. For example, specify no cream backgrounds rather than no generic AI look.

Effort Levels: Why Medium Is the New High

The biggest practical change for developers is the effort parameter. Opus 5.5 defaults to medium effort — and at medium, it matches or beats Opus 5 at high effort on real benchmarks.

On CursorBench 4.0, Opus 5.5 at default (medium, 52.5%) beats Opus 5 at max (46.6%) for roughly one‑fifth the cost. On FrontierCode, Opus 5.5 at medium scores 54.6% versus 54.4% at max — medium is nearly at the ceiling, Anthropic's launch page shows.

Higher effort levels show diminishing returns and are prone to overthinking. Anthropic partners report that medium often matches higher settings. The recommendation is to start at medium and only raise effort when you have a specific failure case that higher effort fixes.

The Counterintuitive Cost Story: Opus at Medium vs. Sonnet at Max

On a per‑token basis, Opus 5.5 costs exactly double Sonnet 5.5 ($4/$20 vs $2/$10 per million input/output tokens). But per task, the math flips because Sonnet at high effort writes dramatically more tokens and Opus finishes in fewer steps.

Real per‑task data from Artificial Analysis via Lindy.ai shows the effect clearly:

Sonnet 5.5 at Max effort scores 56 at $5.46 per task. Opus 5.5 at Extra High scores the same 56 at $3.46 — $2 less per task. And Opus 5.5 at Medium ($1.34) lands within one point of Sonnet at Extra High ($2.01) while costing 33% less.

As 1 put it: "Anyone who regularly cranks Sonnet up to 'Extra High' or 'Max' effort ends up paying about as much as with Opus 5.5." Anthropic says the more capable model finishing in fewer steps partially offsets the higher per‑token price.

Five Gotchas the Headlines Don't Cover

Beyond the headline advice, there are several edge cases that developers will hit in production:

1. Multi‑turn chat re‑examination latency. Opus 5.5 may go back over earlier answers while thinking about a new message — even a short follow‑up adds hidden thinking time. Anthropic recommends adding to Q&A chat apps: "Once you have answered something, treat that answer as done — don't go back over it," per ClaudeFA.ST.

2. Agent continuation loops. Opus 5.5 sends progress updates that end turns with text instead of tool calls. Systems that treat every text‑only turn as "finished" will stop prematurely. The fix: maintain a TASKS.md checklist, send continuation messages when items remain open, and cap at 2–3 automatic continuations.

3. Per‑message thinking nudges still work. Anthropic's docs don't say prompt wording has zero effect — only that effort is the calibrated lever. Adding "Please think hard before responding" to a specific turn still encourages deeper reasoning, Alpesh Kumar notes in his developer analysis. The recommendation is to remove standing instructions only.

4. Cache invalidation on effort change. Changing top‑level effort between API requests invalidates cached messages entirely. A beta per‑message effort change mechanism exists to preserve cache when switching effort mid‑conversation.

5. max_tokens must account for hidden thinking. Thinking tokens count toward the output token limit even when invisible (default display: "omitted"). Limits sized for Opus 5's no‑thinking integration will cut Opus 5.5 replies short. Anthropic recommends 128,000 for long agentic coding turns and 64k minimum for xhigh/max.

The Contradiction: Anthropic Taught You Those Habits

There's an irony at the heart of this guidance. Anthropic's own documentation from the Opus 4.7 era told developers to add Anthropic's platform docs from the Opus 4.7 era told developers to add "This task involves multistep reasoning. Think carefully before responding" The recommended CLAUDE.md templates included "Always think step by step" instructions Now the company is telling developers those same instructions are dead weight.

This isn't a bug — it's a consequence of model architecture evolving faster than prompting best practices. Opus 5.5's adaptive reasoning represents a genuine architectural shift: the model doesn't need to be told to think because it's always thinking. The prompt folklore that accumulated over 18 months of Claude development needs a reset.

For builders, the takeaway is practical: audit your CLAUDE.md, strip the thinking instructions, set effort to medium, switch to single‑message goal‑setting, and attach visuals directly. The model got smarter — your prompts need to get simpler.

Sources

  1. 1.THE DECODER(the-decoder.com)
  2. 2.Search Engine Journal(searchenginejournal.com)
  3. 3.UX Planet(uxplanet.org)
  4. 4.Alpesh Kumar(alpeshkumar.com)

Share this article

PostShare

More on This Story

Related News

Anthropic Wants You to Build on Claude — Then It Competes With You

Nov 15, 2026

Anthropic Wants You to Build on Claude — Then It Competes With You

Anthropic is aggressively expanding into enterprise software with Claude Sheets, Whiteboards, and industry-specific tools, raising alarms among developers who say the company is following Amazon's playbook: watch what sells, then ship a competing version.

anthropicclaudeclaude-enterprise
Anthropic Launches Claude for Manufacturing With 17 Factory Skills

Nov 14, 2026

Anthropic Launches Claude for Manufacturing With 17 Factory Skills

Anthropic launched Claude for Manufacturing, a plugin packing 17 skills and 4 connectors to Autodesk Fusion, Infor, Microsoft 365, and Oracle NetSuite. The AI drafts quotes from live inventory, diagnoses breakdowns, and compares supplier options — but always stops for human approval before anything reaches a customer.

claude-for-manufacturinganthropicclaude
Anthropic IPO 6x Oversubscribed at $1.4T Valuation, Trading Nov 19

Nov 13, 2026

Anthropic IPO 6x Oversubscribed at $1.4T Valuation, Trading Nov 19

Anthropic's IPO is six times oversubscribed at a $1.4 trillion valuation, with pricing set at $70–$78 per share for an 825 million share offering ahead of a November 19 Nasdaq debut under ticker ANT. The company simultaneously launched Claude for Manufacturing — a 17-skill plugin for factory-floor workers — and Claude Meetings, but a shadow hangs over Nvidia's $10 billion private-placement pledge following Jensen Huang's resignation as CEO. For builders, this is the most consequential AI IPO since SpaceX: it will set the pricing benchmark for AI companies and determine how much capital flows into Claude's developer ecosystem.

anthropicipoanthropic-ipo