{"slug": "12-prompt-habits-to-change-for-claude-opus-5-5", "title": "12 Prompt Habits to Change for Claude Opus 5.5", "summary": "A developer published a set of twelve prompt-writing habits to change for Anthropic's Claude Opus 5.5, released September 22 at $4 per million input tokens and $20 per million output tokens, roughly 20% cheaper than Opus 5. The guidance argues that older prompting conventions — such as \"think step by step\" lines and defaulting to high effort — now work against the model, and recommends defining a finish line in the first message, starting at medium effort, and naming the specific stops and design styles to avoid.", "body_md": "Claude Opus 5.5 came out on September 22. It costs $4 per million input tokens and $20 per million output tokens, 20% less than Opus 5. Anthropic says it writes output more than 30% faster and usually finishes the same task with fewer tokens.\n\nThose numbers get the attention. The change I care more about is quieter: several prompt habits we built for older models now work against this one.\n\nHere is how I think about it. Treat Opus 5.5 like a senior contractor. You don't tell a good contractor to \"think hard.\" You tell them what the finished kitchen looks like, what they must not touch, and when to call you. Every habit below is a version of that idea.\n\n*The 12 habits at a glance*\n\n**1. Define \"done\" in the first message.**\n\nOpus 5.5 is built for long, multi-step work. Anthropic says it holds up on multi-hour audits and migrations better than Opus 5. A long run needs a finish line, though. Without one, the model decides for itself when it has done enough, and it may stop early to report.\n\nSo I give the whole task in one message, say what done means, and say when it should stop and ask:\n\nMigrate the payment endpoints from the old client to the new one. Done means: every endpoint uses the new client, the old client is deleted, and the test suite passes. Stop and ask me only if a test fails for a reason you can't explain.\n\nThat prompt has three parts: the task, the finish line, and the one case where the model should stop.\n\n*Source: claude.dev*\n\n**2. Delete \"think carefully\" lines.**\n\nOpus 5.5 always thinks before it replies. You can't turn thinking off on this model; the API returns an error if you try. The model decides how much to think, and the effort setting controls it.\n\nThat makes \"think step by step\" dead weight. Anthropic tested this in a chat product and found that removing a \"think carefully\" line made replies start sooner with no clear drop in quality. If I want a quick answer, I say \"Answer directly.\" If I want more or less thinking, I change the effort level.\n\n**3. Start at medium effort.**\n\nMedium is the default on Opus 5.5. Opus 5 defaulted to high. This isn't a downgrade, because effort names don't mean the same amount of thinking on every model. Anthropic reports that Opus 5.5 at medium matches or beats Opus 5 at high on coding and knowledge work, and that low comes close on several coding tests at much lower cost.\n\nThere is a catch if you copy your old setting over. At the same level, Opus 5.5 thinks more per turn than Opus 5, especially at xhigh and max, so expect longer turns and bigger bills. My rule is simple: don't pay for xhigh or max until you've measured that it helps.\n\n**4. Name the design styles you don't want.**\n\nAsk Opus 5.5 for a page with no design direction and it falls back on a few default looks. \"Avoid a generic AI look\" doesn't fix that. It mostly swaps one default for another.\n\nA list of specific patterns works much better. For example, tell it not to use a cream or off-white background, italic accent words in headlines, numbered \"01 / 02 / 03\" section labels, monospace labels, or pill-shaped buttons. Then look at what it picked instead. If you don't like that either, add it to the list and ask again.\n\n**5. Tell it which stops you want.**\n\nOpus 5.5 keeps you posted as it works. On long tasks, some of those updates end the turn: a summary that names the next step without taking it, an offer to continue, or a list of decisions that don't block anything. If you aren't watching, the run just sits there.\n\nThe fix is to name the stops. A short rule in CLAUDE.md does it:\n\nWhen a step doesn't need my input, keep going. Put status notes in the same message as your next action. Stop and ask only when you can't continue without me, or before anything destructive: deleting data, force-pushing, or changing anything outside this repository.\n\nKeep that last sentence. A rule to keep going means fewer check-ins, so you still want a hard stop before anything risky or hard to undo, and permission prompts on for destructive commands.\n\nIf you pair-program and want the opposite, ask for a one-line plan before it starts and a short recap at the end. It follows either style.\n\n**6. Keep the task list in a file.**\n\nLong runs fill the context window, and Claude Code then summarizes older turns. A checklist in a file survives that summary. It also shows you where the run is without scrolling back.\n\nKeep a checklist in TASKS.md. Tick each item when it's done, and add anything new you find.\n\n**7. Ask for subagents by name.**\n\nFor an audit, a migration, or a review across a large codebase, ask it to split the work across subagents and check each result:\n\nAudit every service in services/ for the retry bug in the linked issue. Give each service to its own subagent. When a subagent reports back, check its evidence before you accept it.\n\nIn my view, it's better to ask for this outright than to hope the model decides on its own that the task splits well.\n\n**8. Steer instead of restarting.**\n\nIf you remember something mid-run, type it while the model works. Runs are longer now, so a restart throws away more work. In Claude Code you type the message and press Enter, for example \"Also keep the old endpoint names as aliases.\"\n\n**9. Ask what it couldn't confirm.**\n\nIn research and analysis, \"I couldn't find this\" is some of the most useful text in the answer. Ask for it and it's easy to spot:\n\nMark anything you couldn't confirm, and say where you looked.\n\nWhen a long run ends, read what it needs from you first, such as an open decision or a change waiting for approval. Anthropic says Opus 5.5 reports more plainly than Opus 5 on what it did, what it found, and what it needs, so the summary is worth your time.\n\n**10. Review before a person does.**\n\nHave Opus 5.5 review a diff or pull request before a teammate sees it. Early testers reported that it catches more bugs than Opus 5 with fewer false alarms. Keep the request narrow:\n\nReview the diff on this branch against main. List only problems you'd block the merge for. For each one, give the file and line, why it's wrong, and how to show it fails.\n\nFor important changes, I'd also get a second review from a different model family. Different models miss different things.\n\n**11. Don't ask it to show its reasoning.**\n\nThis one surprised me. Opus 5.5 runs a safety check for \"reasoning extraction,\" and a prompt that asks it to reproduce its internal reasoning in the reply can be refused. If older prompts said \"show your reasoning\" to stand in for thinking, remove that line.\n\nYou can still get the why. Ask for it as a normal answer: \"Explain why you chose this approach in three sentences.\"\n\n**12. Attach the screenshot. Don't retype the numbers.**\n\nOpus 5.5 reads charts, diagrams, and screenshots more accurately than Opus 5. Anthropic says that even at its lowest effort it read values off dense charts better than Opus 5 did at its highest. It's also better where meaning depends on position: which boxes an arrow connects, what changed between two versions of a diagram, or when a meeting starts in a calendar screenshot.\n\nSo attach the image and ask a specific question. Retyping the numbers just gives you a chance to get one wrong.\n\nMost of these habits remove instructions we wrote to make up for older models: \"think harder,\" \"show your work,\" the retyped chart. The rest add what the model can't guess: the finish line, the stops you want, the styles you dislike.\n\nIf I had to fit it in one line: say what you want, what done means, and when to stop. Then let it work, and read what it needs from you when it's done.", "url": "https://wpnews.pro/news/12-prompt-habits-to-change-for-claude-opus-5-5", "canonical_source": "https://dev.to/shreyshah/12-prompt-habits-to-change-for-claude-opus-55-44m2", "published_at": "2026-09-25 22:40:24+00:00", "updated_at": "2026-09-25 23:30:05.715291+00:00", "lang": "en", "topics": ["large-language-models", "ai-tools", "generative-ai", "ai-products"], "entities": ["Anthropic", "Claude Opus 5.5", "Claude Opus 5"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/12-prompt-habits-to-change-for-claude-opus-5-5", "markdown": "https://wpnews.pro/news/12-prompt-habits-to-change-for-claude-opus-5-5.md", "text": "https://wpnews.pro/news/12-prompt-habits-to-change-for-claude-opus-5-5.txt", "jsonld": "https://wpnews.pro/news/12-prompt-habits-to-change-for-claude-opus-5-5.jsonld"}}