What Claude is genuinely good at, judged by what people produce

What can Claude AI do on a wet Wednesday, as opposed to in a demonstration? The most useful answer available comes from Anthropic's own Economic Index, which counts what people produce in their conversations. That is a better guide than any feature list, because it records behaviour rather than intent.

What people make with it

In work conversations, documents are the largest single output at 20%. Explanations account for 9%, drafted emails for 7% and analyses for 6%. None of that is exotic. It is memos, minutes, briefs and the reply you owe Legal: the paperwork an organisation runs on, and nothing anyone would put in a demo.

The work it takes off your desk

Three jobs come up in nearly every session I run, and people are usually surprised to hear they are allowed to hand any of them over.

• The first draft of anything nobody enjoys starting: the supplier review open in a tab since March, the policy untouched since 2021, the note to the board explaining a delay. Claude will not write the version you send. It removes the blank page, which is the expensive part.
• Long material turned into something you can act on. A hundred-page consultation response, a stack of tender documents, the transcript of a two-hour meeting nobody minuted. Ask for the five decisions and who is on the hook for each.
• Formatting and reconciliation. Inconsistent job titles into one convention, dates from forty emails into a single timeline, a spreadsheet checked against the contract it was built from. Dull, error-prone, and done badly by tired people at five o'clock on a Friday.

Translation runs both ways, and nobody asks upward

Of everything on this page, translation is the strength people under-use most. Legal advice rewritten so an operations lead can act on it is the request I hear constantly. The other direction is asked for almost never: an operational reality written up so a lawyer can rule on it, an engineering constraint set out so a customer will accept it, a set of accounts turned into three sentences for a trustee who does not read accounts.

The usage figures hint at why. Across all conversations, explanation is the single largest thing people produce, ahead of documents. At work that inverts, with documents first and explanations dropping under a tenth. People explain things to themselves in private and stop the moment they are on company time, which is the wrong way round, because at work the person who needs the explanation is somebody else.

Knowing this half of it matters less than knowing what it is bad at before that costs you something. The hours it frees come back in particular places, not across the board.

The parts that take practice

Two habits separate the people who get steady work out of it from the people who get the occasional good afternoon.

One is saying what a summary is for. Ask for a summary and you get an even shading of everything, each section weighted the same, which is the least useful document in the world. Name the decision you have to make on Thursday and the same material comes back sorted by what bears on it. There is a floor to this: under about ten pages you will read the thing faster than you can specify what you want from it.

The other is supplying the version that failed. Everyone thinks to paste in a good example. Far fewer paste in the draft that came back covered in comments and say what the comments objected to, which is how you describe the thing to avoid, and avoidance is the harder half of any brief.

So try the failed-draft trick this afternoon, on the last thing of yours that came back marked up. It takes four minutes, nobody demonstrates it, and it is the single habit I have watched change how senior people write.