Claude

What it is

Claude is the language model almost everything here gets built with. It reads text, writes text, and can operate tools you hand it. That sounds small, and it is the reason so many processes became possible in the last two years that used to hang on a person: anywhere someone had to read something, place it, and phrase a reply, a model can now stand.

The limit behind it matters. A model is not a program that does the same thing every time. It is a very good suggestion engine.

Where it holds

Anything that is language and where a good approximation is enough: product copy, draft replies, summaries of long threads, sorting enquiries into categories, translating between two formats nobody documented.

The second, badly underrated place is the workbench. Claude Code sits next to the editor and can read a whole project before it changes anything. For what I build that is now the bigger lever: not the text the model writes, but the system that gets built faster with its help.

Where it bites

It does not decide reliably. The same question asked twice can give two answers. For a draft that is irrelevant; for a rule it is the opposite of usable. So the line here is absolute: rules decide, the model phrases. A model expected to remember the policy is a flaw in the blueprint.

It invents convincingly. A wrong number sounds exactly as confident as a right one. Anything that has to be verifiable therefore enters the process from a database and is never recalled by the model.

It costs per word, and the window is finite. The more you push in, the more expensive it gets and the worse the answers become. The real craft is not the prompt but the selection: what exactly does this one step get to see.

What it looks like when it runs

The answer to a recurring question is already drafted, pulled from your own data rather than from what a model assumes, and a human releases it before it goes out. The same goes for product copy, summaries, and everything somebody types on the side today. The gain is not the text itself. It is the half hour nobody spends on a first draft any more, and the same shape of answer across hundreds of cases.

What to think about before choosing it

The question is rarely which model. The question is which decision you are willing to hand over. As long as the answer is “none”, a model is still useful, it just writes drafts. The moment something goes out unchecked, you need rules first, an escalation for “no rule fits”, and someone who notices when quality drifts.