AI recommendations are about to get uncomfortably good. According to a report from The Information, Anthropic is building a new Claude model that's noticeably better at working out what a user actually wants — even when the user hasn't explained it properly — and at suggesting off-script solutions instead of the same canned advice every chatbot hands you.
That sounds like a small upgrade. It isn't. Most of us are terrible at asking for what we want, and every AI tool you've used so far has been quietly punished for that.
What happened
The Information reported that Anthropic is developing a Claude model with sharper "intent recognition" — the ability to figure out the goal behind a request rather than just the words in it. A source who had already tried the model told the publication two things: it reads intent better, and it's much stronger at finding unusual solutions instead of the standard template answer.
The story travelled fast. The Telegram post where I first saw it pulled 1,372 views and 45 shares, which for a niche AI channel is a lot of forwarding.
The reporting has since been picked up more widely, including in this ForkLog write-up, and it lines up with Anthropic's own published thinking about models that keep improving themselves and about how Claude is supposed to behave — see the company's recursive self-improvement notes and its constitution.
One caveat before you get excited: the name. Some versions of the story are circulating as "Claude 3.5 Turbo." Anthropic hasn't announced any model by that name, and the leaked details don't confirm one. Treat the label as noise and the behaviour as the interesting part.
Example one: you ask for a bug fix, you get an overhaul
The report's first example is coding. You ask the AI to fix one specific bug. Instead of patching a few lines, the model works out the whole class of problems that produced the error, finds the shared root cause, and proposes killing that entire family of bugs at once.
You asked for a working function. The model decided you actually want a reliable system.
Think of it like a plumber. One plumber fixes the leak under the sink and leaves. Another looks at the pipe, sees the whole run is corroded, and tells you the leak will be back next winter. The second one is more useful — and more annoying, because the bill is bigger.
Example two: the investing question
The second example is investing. You ask how to invest in stocks, and instead of the usual lecture about P/E ratios, financial statements and diversification, the model reportedly goes deeper into what you're really trying to achieve. The published excerpt cuts off mid-sentence there, but the point is clear: it doesn't stop at boilerplate.
Why "reading intent" is a bigger deal than it sounds
Intent is just the goal behind the request. When you tell a pharmacist "my head hurts," you don't want a lecture on headaches — you want something that stops the pain and won't knock you out before a meeting. A good pharmacist works that out from context. A bad one reads the label on the box.
Until now, AI has mostly been the box-reader. It answers the sentence you typed. That's why prompts became a skill, and why so many people got mediocre results: they said "fix this bug" when they meant "make this system stop breaking," and the model politely did the literal thing.
A model that infers the goal skips that step. It behaves like a good contractor who walks the site before quoting, rather than someone who does exactly what the drawing says and leaves you with a roof that leaks.
What it means for you
You don't need to be a developer for this to change your day. Here's what "it guesses what you actually want" looks like in ordinary life. If you want free tools to practise the skill of asking better questions, there's a useful free collection at MyKreaTool — worth a browse before you upgrade anything.
At home
You type: "my kitchen tap drips, what part do I buy?" A literal assistant names a washer. A goal-reading one works out that you're a renter on a Sunday with no car, and tells you which fix you can do with a spanner and which one you should hand to the landlord. Same question, completely different usefulness.
At work
You ask for help writing a weekly status update. A template model produces bullet points about tasks. The useful version figures out that your real problem is that your manager thinks the project is stalling, and it builds the update around evidence that it isn't. You asked for formatting. You needed political cover.
Running a small business
You ask whether to raise prices. The generic answer is "test it." The better answer digs into your actual constraint — maybe it's not price at all, it's that your cheapest customers soak up most of your support time. That's the model spotting the goal you didn't state, and it's where this stuff starts to pay for itself.
Studying
You paste a chapter and say "summarise this." A literal tool gives you a shorter version of the chapter. A goal-aware one works out you're three days from an exam and quizzes you instead, or explains the one concept everything else in the chapter depends on. You asked for a summary because you couldn't name what you were missing.
Creative work
You send a brief that says "make it punchier." A template model adds exclamation marks. A model reading intent notices the copy is aimed at people who are already sold and suggests you stop selling and start explaining. That's a rewrite you'd never have asked for and probably needed.
Earning money
Freelancers live on this. You ask how to price a logo project; the boilerplate says "charge by the hour." The intent-aware answer asks what you're actually optimising for — cash this month, a portfolio piece, or getting off a bad client — because those three answers point in opposite directions.
How to try it right now
You can't get the unreleased model yet. But you can train the same behaviour out of what already exists, starting free.
1. Open Claude on the free plan at claude.ai. No card, no setup — this is the same tool family the report is about, just an older model.
2. Give it a deliberately vague request. Something real from today: "help me with my CV," "my site is slow," "I need a content plan." Don't tidy it up.
3. Ask the meta-question. Type: "Before you answer, tell me what you think my real goal is here, and what I probably left out." This single line is the closest thing you have to the rumoured feature.
4. Ask for the class, not the symptom. If it's a bug or a broken process, say: "Don't just fix this one case — what family of problems is this, and what's the shared cause?" You'll get the coding example from the report, using a model that shipped months ago.
5. Push back once. Say "that's the textbook answer — what would you do if you knew my actual constraint?" Then name the constraint. This is the step almost everyone skips.
6. Save the prompts that worked. Three good ones beat fifty tabs.
Upsides and what changes
Templates stop being the product. If a model can infer your goal, the value moves from "here's the standard advice" to "here's what you should have asked." That's a real shift for anyone who's spent hours fighting with prompt wording.
It also compresses expertise. The gap between a beginner's question and an expert's question is mostly knowing what to ask for. A model that fills in that gap gives beginners expert-shaped answers — which is genuinely good news for freelancers, small teams and anyone without a specialist on call.
And for coding specifically, catching a family of bugs instead of one instance saves the most expensive kind of time: the time you spend finding the same problem again next month.
Limitations
The honest part. This is a leak, not a launch — nobody outside Anthropic has benchmarked it, and "the model understands what you really want" is exactly the kind of claim that sounds amazing in a demo and wobbly in week three. Guessing intent means guessing, and a confident wrong guess is worse than a dumb literal answer, because you'll trust it. The bug-fix example is the perfect illustration: "I'll rewrite the whole function" is wonderful when it's right and a disaster when it isn't — half your codebase changes, one thing gets ten times faster, and something unrelated quietly breaks. On top of that, a model that infers goals is a model that infers things you didn't mean to share, so keep sensitive data out of your prompts until you've seen how the released version handles it.
Conclusion
The pitch is simple: AI stops answering the question you typed and starts answering the question you meant. That's a bigger jump than another point on a benchmark, and it's coming from a company that has already published its thinking on models that improve themselves.
One thing to do today: go back to the last request you typed into any chatbot, paste it in again, and add one line on top — "Tell me what you think my real goal is before you answer." You'll learn more about prompting in one minute than in an hour of reading about it.
Comments 0