How to Prompt Claude Opus 5 (In Plain English)

Short answer: Claude Opus 5 is a smart model, but it picked up a few new habits that trip people up. Once you know them, your results get noticeably better. The plain version: let it think, stop cranking every setting to the max, tell it to keep answers short, and delete any instruction that tells it to double-check its own work. Here is the whole thing in plain English, with lines you can copy straight into your prompt.

If you have used Claude Opus 5 and felt like the answers were a little long, or it did more than you asked, you were not imagining it. The model behaves differently from older ones, and a prompt written for the old version leaves a lot on the table. None of the fixes are complicated.

The basics that always work

Before the Opus 5 specifics, four things that make any model better. Think of Claude as a sharp new hire who is great at the work but knows nothing about your project.

• Be specific. "Write a launch email" gets you a generic email. "Write a launch email to existing customers, friendly but short, one clear call to action" gets you something usable.
• Say why. If you tell it the reason behind a request, it makes better choices on everything you did not spell out. "Keep it simple, this is for non-technical readers" goes a long way.
• Show an example. One or two examples of what good looks like beats a paragraph of description.
• Give it a role. Starting with "You are a senior copywriter" or "You are a careful code reviewer" focuses the whole answer.

That is the foundation. Now the part that is specific to Opus 5.

The Opus 5 habits worth knowing

This is drawn from Anthropic's own guidance for Opus 5, plus what I have seen using it day to day. Here is the short version before the detail:

It thinks by default now, so let it

Older models did not "think" unless you asked. Opus 5 does, automatically. That is a good thing, it makes answers better. The only catch is that its thinking and its answer share the same length budget, so if you set a tight limit on response length, it can now run out of room and cut off mid-answer. Give it more headroom and you are fine.

Do not crank the effort dial to the max

Opus 5 has an "effort" setting that goes from low up to max. The instinct is to max it out for the best answer. Skip that instinct. Anthropic's own advice is to start around the middle and only reach for the top setting on genuinely hard coding or agent work. On this model, low and medium are surprisingly strong, and they are cheaper and faster. Higher effort is not automatically better, and it often just costs you more tokens and time for the same result.

It talks a lot, so tell it to be brief

Opus 5 writes longer answers than older models by default. Here is the part that surprises people: turning the effort setting down does not reliably make the answers shorter. The thing that works is simply asking. A short instruction to be concise cut response length by around 20% in Anthropic's testing. Drop this into your prompt:

Keep your answers short and focused. Put almost all of the response on the actual answer, and keep any caveats brief. When I ask you to explain something, give the short version unless I ask for more.

Stop telling it to double-check its work

This is the one that catches everyone, because it is the opposite of the usual advice. For years, "double-check your answer before you finish" made models better. On Opus 5 it backfires. The model already verifies its own work without being asked, so a "double-check" instruction makes it check twice, which just burns time and money for no gain. If you have that line in your prompt, delete it. You will get the same quality, faster.

Give it the whole task up front

Opus 5 is at its best when you hand it the full task in one clear brief and let it run, rather than dripping instructions out one message at a time. Front-load everything: the goal, the constraints, and what "done" looks like. It holds the whole picture well, so give it the whole picture.

Keep it on the task you actually asked for

Left to its own devices, Opus 5 will sometimes do more than you asked, or quietly reshape the task into what it thinks you meant. Usually helpful, occasionally not. When you want it to stay in a tight lane, say so:

Do exactly what I asked, nothing extra. Make small judgment calls yourself, but check with me before changing the scope or doing anything hard to undo. Finish the whole task, and if you cannot finish part of it, tell me plainly what is left.

If you build with agents, rein in the helpers

This one is only for people wiring Opus 5 into a tool or agent setup, not for chat. Opus 5 likes to spin up helper agents ("subagents") to split up work. That is great for genuinely big, independent jobs and wasteful for small ones, because every helper re-reads the context and reports back. If cost matters, cap it:

Only split work across helper agents when the pieces are genuinely large and independent. Do not spawn a helper for something you could finish in a few steps yourself, and never use one just to double-check your own work.

The one-minute version

If you remember nothing else: let it think, keep the effort setting modest, ask for brevity in plain words, and take out any "double-check yourself" line you copied from an older guide. Those four changes alone will make Opus 5 feel sharper and cost you less.

Prompting is a real part of whether an AI feature works in production or quietly wastes money, and it is most of what separates "the demo was cool" from "this actually ships." That is the kind of thing I help founders get right when we build AI into a product. If you have an AI feature in mind and want it done properly, see what I work on or book a call.