You open Claude Opus 5 and type it out loud as you go: "First do step one, then step two, then step three, then step four." Every stage nailed down tight — that feels like the safe way to do it. You wait a while. It finally comes back — with a pile of extra stuff you never asked for, and your token bill just ticked up another notch.
That exact moment is the number-one habit Anthropic's own prompt engineers are trying to break.
The old habit of chopping tasks into "1, 2, 3, 4" is now working against you
Ben, a three-time founder who runs a large AI community and an AI agency (channel: Ben AI), went through Anthropic's official prompting guides for Claude Opus 5 and Fable 5, along with the internal discussion around them, and landed on one conclusion: most people are still prompting the new models the way they prompted the old ones — and this generation runs on fundamentally different logic.
Boris Cherny, the creator of Claude Code, called out this exact habit directly in a recent keynote at Y Combinator: a common mistake is giving Claude overly precise instructions, insisting it follow one specific step after another. For this generation of models, that isn't the right move. Instead, you should step back: describe the task, state the constraints and the boundaries, explain what "done" looks like, then let the model run on its own — and come back later to check the result. He said flatly that this approach didn't work six months ago. Today, it works remarkably well.
Why Anthropic says "say the whole thing up front" is the new fix
The reason isn't complicated: Claude Opus 5 and Fable 5 are trained to handle tasks end-to-end — done in one continuous pass, not obediently ticking through a pre-written checklist. Anthropic's own prompting best-practices documentation says as much, plainly: for complex tasks, hand over the goal, the reasoning behind it, the constraints, and what "done" looks like all at once, and let the model decide how to break the work down itself. That tends to outperform locking down every single step in advance.
That's the exact framework Ben used to prepare this video: give the model a broad goal and direction, explain why it matters, what the constraints are, and what the finished product should look like — then let it run for ten-odd minutes and come back with a usable draft outline. He also flags the flip side: small tasks don't need the full framework every time. Pulling out that much scaffolding for a five-minute job is overkill.
Even Anthropic gets "interviewed" before it starts
The problem is, complex tasks rarely arrive fully formed in your head the moment you sit down to write the prompt. Ben mentions that Anthropic's own team leans on something they call "Interview me" — the model turns the questioning around on you, asking a string of questions to surface the parts of the task you hadn't fully worked out yet, then compiles a task brief, or even generates a full prompt from scratch. Ben used it to plan his own personal data dashboard: he wasn't sure what he wanted going in, and only after letting the model "interview" him did he end up with a task brief he could actually act on. His takeaway: spending more time up front planning, rather than writing fast and revising endlessly, ends up saving both tokens and time with a model this capable.
Another pre-work step he uses, borrowed from Andrej Karpathy's advice, is talking through the task out loud into a speech-to-text tool for a few minutes — that tends to hand Claude more context than typing does. But spoken thoughts ramble, so Ben built himself a small tool that turns that messy dictation into a structured prompt — his own build, not an Anthropic feature.
Telling Claude "what" without "why" reads like a memo with no context — say "why"
Anthropic's guide is explicit: when the model understands why you're asking for something, performance on complex tasks improves noticeably. That's because complex tasks are full of small judgment calls you never spelled out in advance — and once the model has the bigger picture and the underlying goal, it can make those calls in a way that actually fits. Anthropic even offers a template sentence for this: state clearly who the work is for, and what that person actually needs to get out of the result.
Ben's example is adding one line of "why" to the same prompt — explaining that he's making a video about how prompting Claude has changed, aimed at non-technical professionals and business owners who want to automate their work with Claude, and that they need more than theory — they need techniques and frameworks they can actually use. It's the same logic as delegating to a person: telling someone just "what" to do is faster to write, but the delegation that actually works is the one that includes "why."
A task with no finish line burns your tokens, not its own
Another thing Boris called out specifically in his YC talk was the "finish line." Claude 5 models tend to run longer and lean toward doing too much rather than too little — without a clear signal for when the task is done, it can keep going, and burn through a meaningful chunk of tokens doing it. There are two parts to fixing this: describe what the final output should look like and what tone or style it should be in, and give a reference example to check against. Ben writes the output spec directly into his prompt — a fixed number of suggestions, each backed by a concrete example and sourced specifically to something Anthropic has officially endorsed. He also flags a common Opus 5 quirk: it tends to talk more than it needs to, and a single line — "output this in plain, simple language" — corrects it.
"Don't do X" and "don't do X because Y" land very differently with Claude 5
Plenty of people default to hard bans — "absolutely do not." But Anthropic's newer research on context engineering points out that this generation of models responds noticeably better when it also knows the reasoning behind an instruction. The constraint still needs to be there — just don't hand it over as a bare "you can't do this." Always follow it with why. Ben's example: rewriting "don't include any claim Anthropic hasn't backed" into "make sure every point listed is backed by Anthropic's own team" — the difference being that now you're actually holding evidence, not just issuing a ban.
You might be paying Claude to do something it already does by default
A lot of people are used to tacking on "double-check your answer before responding," or having a sub-agent re-verify the output, as a way to cut down on hallucination. But Ben notes that Opus 5 and Fable 5 are already trained to automatically re-check, verify, and correct their own output — asking for one more check on top of that just costs more money without improving the result. The same goes for "think step by step," "explain your reasoning," and writing in all caps for emphasis — Anthropic has observed that these can actually cause this generation of models to overreact. That applies to your CLAUDE.md files and skill files too — according to Ben, this is exactly why Anthropic built a tool called "Rule Rewriter": to take hard-coded rules and rewrite them into guidance that carries its own reasoning.
Opus 5 sounds like a memo? One line fixes that
A lot of people who use Opus 5 have run into this: the word choice is dense and stiff, and the output is hard to read. Ben notes that Anthropic's suggested fix is to write the tone instruction directly into your CLAUDE.md or system prompt — explicitly asking for output that's "focused, concise, on-point, and free of jargon and filler."
String these seven things together, and they're really pointing at the same idea: with this generation of models, the move isn't memorizing more incantations — it's shifting how you hand off a task in the first place. Treat Claude like a smart new hire who just started and doesn't yet know how you like things done. Instead of only telling it "what," tell it "why" too, and work out "what counts as done" before you even start typing. Next time you open Claude Opus 5, maybe put away the "step one, step two, step three, step four," swap in one full description of the task and a clear finish line — then go make a coffee, and come back in ten minutes to see what you've got.
FAQ
Q1: Why doesn't the old "step by step" prompting style work anymore with Claude Opus 5 and Fable 5? Because this generation of models is trained to handle tasks end-to-end. Chopping a task into overly granular steps actually limits the model's ability to make its own judgment calls, which tends to produce output that's bloated and off-target — and it burns unnecessary tokens along the way.
Q2: What is the "Interview me" approach Anthropic mentions? According to Ben's video, this is an approach Anthropic's own team reportedly uses internally: before handing a complex task to a strong model, they let the model turn the questioning around and ask a series of clarifying questions first, surfacing details you hadn't worked out yet, then produce a task brief or a ready-to-use prompt. This specific name and its exact mechanics currently come only from Ben's account and haven't been individually verified against the official pages cited in this article.
Q3: Why does it matter so much to tell Claude when a task counts as "done"? Because Claude 5 models tend to do more than expected and run longer than expected. Without a clear completion standard, the model is likely to keep expanding the task on its own — which drives up token usage along with it.
Q4: Do I still need to ask Claude to double-check its own answer? For Opus 5 and Fable 5 specifically, Anthropic's own observation is that these models already have automatic re-checking and self-correction built in. Asking for an extra verification pass on top of that mostly just adds cost without improving the result.
Q5: Opus 5's writing feels too formal and jargon-heavy — what fixes it? You can add a single line to your CLAUDE.md, your Claude Desktop global instructions, or your system prompt, asking directly for output that's "focused, concise, and free of jargon and filler." Anthropic's own guide recommends this exact approach for correcting tone.
Q6: Do these prompting habits actually affect AI token costs? Directly, yes. Whether it's skipping a redundant verification step or setting a clear completion standard, both are really about cutting unnecessary token consumption. In other words, whether your prompt is written well or not is itself a form of cost management.
Source Disclosure
This article is adapted from the YouTube video "Anthropic Just Revealed 7 New Rules for Prompting Claude 5 Models" (channel: Ben AI, published 2026-09-02, original link: https://www.youtube.com/watch?v=HDmBwU5uvEE). The video's caption track is in Bengali; this piece was rewritten based on its meaning rather than translated word for word.
The official prompting best practices referenced in this article correspond to the following Anthropic documentation:
Prompting Claude Opus 5 (https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5)
Prompting Claude Fable 5 (https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5)
Claude Prompting Best Practices (https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices)
The specific tool names "Interview me" and "Rule Rewriter" mentioned in this article are personal accounts from the video's creator, Ben AI, describing what he presents as internal or official Anthropic practices. This article has not individually verified their exact naming or mechanics against the official pages above; readers who want to apply them directly should check current official documentation themselves.
Further Reading
You can also read our related pieces on what Claude Fable 5 actually is, how Opus 4.8 compares to Fable 5 on the same real-world build and what that difference costs in tokens, whether the way you write a prompt actually changes your AI token spend, and why prompt caching can sometimes make your AI bill go up instead of down.
Stop Letting the Wrong Prompting Habits Quietly Eat Your Tokens
How you write a prompt doesn't just decide the quality of the answer — it decides how long this call runs, and how much it burns. Instead of waiting for the bill to tell you it's time to change your approach, put the real cost of every call in front of you first. Try AI Token King free → (https://aitokenking.com.tw/home), so your prompting habits can change right alongside numbers that actually add up.
