You Are Still Prompting Like It Is 2025

In late July, a developer publishing under the name zephel01 wrote a short post about a single line of text.

The line had been sitting in his configuration file for a long time. It read: Include a final verification step for non-trivial tasks. Nine words. He had written it at some point, for good reasons, and then stopped seeing it, the way you stop seeing a light switch. It worked. Every run ended with the model going back over its own output.

Then Anthropic published its prompting guide for Claude Opus 5, and one paragraph told him to delete exactly that.

So he deleted it. And something that had been invisible became visible. Removing that one instruction made a whole cluster of related rules in his config pointless. The file got shorter. The behaviour got simpler. Nine words had been quietly holding up a scaffold that no longer had anything to hold up.

Nobody made a mistake here. He added a good instruction, for a model that needed it, and then the model stopped needing it, and the only notice was a paragraph in a document most people skim.

Around the same time, Claire, who writes the How I AI column, was testing Opus 5 against six other models. It came out on top. She scored it 78, ahead of Sonnet 5 at 77 and GPT-5.6 Sol at 76, with straight fives on front-end design. The best result in the set.

She also found it exhausting.

At one point it refused to resolve a one-line merge conflict, on the grounds that the code belonged to someone else's branch. She kept typing variations of just do it. When she asked it to rebuild a benchmark website, she had to tell it to start over, because the actual work was buried under so much commentary about the work. Her verdict on the best-scoring model she tested was that it was brilliant, and annoying.

Hold on to that. It matters later.

Two things happened at once

Opus 5 got better at finishing things. It completes whole tasks instead of leaving stubs. It checks its own output without being asked. It holds instructions steady across a one-million-token context window, which is now both the default and the maximum. It costs exactly what Opus 4.8 cost: five dollars per million tokens in, twenty-five out.

That is the part everyone wrote about.

Here is the part almost nobody did. Three of its improvements arrive at your desk disguised as problems. It writes longer. It expands the job. It hands work to subagents more freely than it used to. Every one of those is a capability that turns into a defect the moment your instructions do not account for it.

Which means the useful way to think about this release is not as a smarter tool. It is as a change in what you are managing.

You have spent two years supervising something that needed supervising. You broke work into steps because it lost the thread. You told it to check itself because it did not. You kept your instructions short because long ones confused it. All of that was correct.

You are now briefing a senior contractor, and you are still talking to an intern. The instructions have not changed. The person receiving them has.

That mismatch produces four things you need to add and one you need to take out.

The brief

The habit to break first is decomposition.

For a long time the right move was to slice a job into stages and feed them one at a time. Look at these images and pull out the branding. Now make a brand guide. Now plan the site. Each step small enough that the model could not wander off.

Opus 5 performs best given the complete task specification up front, and then left alone. Not a stage. The job.

Build a one-page website for Northline Coffee, a small brand that sells coffee beans online. The goal is to help people understand the brand, choose a coffee, and place their first order. Use the logo, colours, product details and photos in the brand asset folder. Build the finished website in the project folder.

Where we start. Where we need to end up. One statement.

The counterintuitive part is what this does to speed. A more specific brief does not slow the run down. In the worked example that prompt came out of, the site was finished in under three minutes, and the model asked no clarifying questions at all. Not because it skipped them. Because a complete brief has nothing left to clarify.

Handing over one instruction at a time is not caution. It is a contractor standing in your driveway, waiting.

The fence

The next behaviour is the one that surprises people who have not read the guide. Opus 5 expands scope. It adds steps nobody requested and applies its own judgement about what the task ought to have been.

So the second thing every prompt needs is a boundary.

Build only the homepage. Link the buy buttons to the shop page that already exists. Keep the current logo, colours and product name. Make small design choices yourself. Only ask me a question if the answer would change the whole website.

Read that last sentence again, because it is doing two jobs and only one of them is obvious.

It stops scope creep. It also stops the opposite failure, which is the model pausing to ask permission for things it should simply decide. That is the failure Claire ran into with the merge conflict. A model that will not touch a one-line change on its own is not being careful. It has not been told which decisions belong to it.

A fence is not a leash. Half of its job is telling the contractor which walls they are allowed to move without calling you.

The shape and the receipt

Now the distinction most people miss, and the most useful idea in this whole subject.

How much it says to you and how big the thing it makes are two different taps. Anthropic's guidance treats them as two separate problems in two separate sections. Conversational verbosity is one. The length of documents it writes to disk is another. You can get a crisp three-bullet reply attached to a bloated forty-section website, and most people who complain about output length are only turning one of the taps.

The shape controls the artefact.

Keep the website to six parts: the top section, three featured coffees, why Northline is different, customer reviews, common questions, and one final buy button. Keep the writing short. The main headline should be eight words or less. Each part should be easy to scan and should have no more than 70 words.

The receipt controls what comes back to you.

When you finish, tell me where the website is and give me no more than three short bullets explaining what you made. Then tell me only what is truly important to know about how it works.

You do not have to be that precise. Six parts, you choose which works. So does someone should be able to read the whole site in ninety seconds. What is not optional is saying something. Left unspecified, the size of the thing is a decision, and the model will make it.

There is a trap here worth naming before you fall into it. You cannot fix long answers by lowering the effort setting. Effort governs how much the model thinks, not how much it says. Turning it down reduces thinking and leaves the visible response roughly where it was. People will try this, spend the saving, and still be reading four paragraphs where they wanted one.

Effort is a deliberation budget, not a quality dial. Setting it to maximum on a routine task is sending a grandmaster into a twenty-minute think on move two. He plays the same move. You just waited for it.

The deletion

Everything so far is something to add. This one is a subtraction, and it is the reason the piece is worth writing.

Stop telling it to double-check.

Opus 5 catches and fixes its own mistakes without prompting. Instructions to verify before responding, re-check the output, or find every possible error do not stack on top of that behaviour. They compound with it. You pay for two inspections and get one inspection's worth of findings. Anthropic's own wording is that removing these instructions reduces wasted tokens with no loss in quality.

That is the nine words in zephel01's config file.

But the line in your prompt is the small version of the problem. The large version is your harness. Verifier subagents. Separate review passes. That step in your pipeline whose only job is to look at the thing that was just produced. Most of that was built to compensate for a weakness, and the weakness has been fixed, and the compensation is still running and still billing.

Keep the standard. Delete the re-check. Telling it what good looks like is still your job. Telling it to look twice is not.

Two more nobody mentions

The five rules doing the rounds stop here. Two things underneath them matter more for anyone running this at scale.

Delegation reversed direction. Opus 5 reaches for subagents more readily than Opus 4.8 did. On genuinely independent, sizeable tracks of work, that pays. On small tasks it multiplies cost and time, because every subagent re-establishes context, re-explores the ground, reports back, and then you read the report. Note which way this flipped. Opus 4.8 under-delegated and needed encouragement. If you wrote that encouragement last year, it is now working against you.

Careful instructions make it look worse. If your code review prompt says to report only high-severity issues, or to be conservative, Opus 5 obeys literally. It investigates just as thoroughly, finds the bugs, and then declines to mention the ones it judges below your bar. Your measured recall drops while its actual bug-finding improves. Ask it to report everything with a confidence and a severity attached, and filter in a second pass.

Both are the same lesson in different clothes. Instructions written to compensate for a weaker model become handicaps on a stronger one.

Back to Claire

Go back to that review and read her complaints as a list.

It was too cautious. It deferred decisions it could have made. It asked when it should have acted. It buried the work under commentary about the work. It produced more text than the task required.

Now read them again as a diagnosis.

No fence, so it asked instead of deciding. No receipt, so it narrated. No shape, so it padded. Every single complaint on that list is a default that three sentences would have removed.

She was not describing a bad model. She was describing an unbriefed one. And she got the highest score in her set anyway, which tells you what the ceiling looks like once the briefing is right.

Try one of these tonight

None of this needs to be taken on trust. Three experiments, twenty minutes each.

Delete your checklist. Take a task you have run before. Run it twice at the same effort. The second time, remove every instruction to verify, re-check, or find all mistakes. Compare the time, the length, and whether the second run actually missed anything. If it did, you have found a case where your standard was doing real work, and that is worth more than following a rule.

Turn both taps. Take a task that produces a file. Run it four ways: no length instructions; low effort and no length instructions; a reply-length instruction; a reply-length instruction and a size instruction. Count the words in the reply and in the file separately. The second run is the one that teaches you the most, because almost nothing will change.

Give it the whole job. Take something you would normally feed in three or four stages. Run it your usual way. Then run it again as a single brief with a fence and a shape, and do not intervene. Count your turns.

If you only run one, run that one. It is the one most likely to change how you work tomorrow.

The part that will not go stale

These rules have a shelf life. There will be an Opus 6, and some of this will reverse again, exactly the way the double-check rule reversed. Anything written today about a specific model is a photograph, not a map.

What survives is the practice underneath. When the model changes, go back through your instructions and ask which ones exist to compensate for something that no longer needs compensating.

Every prompt you have written is a fossil record of some model's weakness. Most of those fossils are still in there, still being read out loud on every single run, still being paid for.

The habit that made you good at this is the habit to check first. Not because it was wrong. Because it worked, and things that work are the last things anyone thinks to remove.