What I look for in an AI tool

When I shipped PromptCraft, one of the side-effects was that people started sending me other AI tools to try. Most weeks it's two or three; some weeks it's a dozen. I now have, statistically, a pretty broad sample.

Patterns have emerged. I want to write down what I actually look at when I evaluate an AI tool, partly because the answer might be useful to other people building, and partly because writing it down forces me to be honest about what I value.

Five traits I trust.

A clear scope. The tool answers, in one sentence, what it's for. Not "an AI assistant for everyone" — that's the absence of a scope. "A grammar checker for Spanish-language academic writing" is a scope. Tools with a clear scope tend to do that scope well, because the maker had to confront tradeoffs instead of waving them away. Tools with no scope tend to do everything 30%.

Defaults that are exposed, not hidden. When the tool runs an AI model, it shows me which model, which version, which parameters. Bonus if I can change them without leaving the happy path. This is rare. Most AI products treat the model as a black box, which signals one of two things: the maker doesn't know, or the maker thinks I can't be trusted to know. Both are bad signs.

Honest limitations. Somewhere in the product copy or docs, the maker has written down what the tool *doesn't* do. Maybe a "won't work for X" note, or a "current limitations" section, or a quiet acknowledgment of a known failure mode. This is the single best signal of a maker I'd trust. People who are honest about the gaps are usually better at the parts that aren't gaps.

Output that respects my time. When I run the tool, I get something I can use, edit, or copy in seconds. Not a flashy preview that requires a paid plan to actually export. Not a multi-step wizard. Not a mystery animation while it pretends to be thinking. Just: input goes in, useful output comes out, friction is minimal.

A real human behind it. Not necessarily a famous one. But somewhere there's a name, a face, a line of writing that sounds like a person rather than a positioning document. The closer the tool feels to "made by a specific human with a specific opinion," the more I trust the product to evolve in ways that aren't just chasing whatever metric the marketing team is measuring this quarter.

Three traits I don't trust.

Vague AI claims that don't survive a closer look. "Powered by AI" is meaningless; "powered by GPT-4" is a starting point; "uses a fine-tuned classifier on top of GPT-4 for [specific task]" is real. The further a product sits from being able to describe what it actually does, the more likely it's a shell with prompt-engineering inside that you could replicate yourself in an evening.

Every feature in the marketing copy. When the product surface lists 40 capabilities, none of them are good. Tools are the result of choices about what to include and what to leave out. A maker who can't tell me which three features matter most is signaling that they don't know either, which means the prioritization on the roadmap is being driven by something other than usefulness.

Slick onboarding that makes me feel watched. I have nothing against a friendly first-run experience. But when the welcome flow asks for my role, my industry, my team size, my use case, my company name, my LinkedIn URL — before showing me what the tool does — I close the tab. I don't owe a stranger a profile in exchange for the chance to find out whether their product is any good.

The reason I write this down is that, building an AI product in 2026, the temptation runs in the opposite direction on every axis. Bigger scope feels safer. Hidden defaults feel more polished. Honest limitations feel like weakness. Anonymity feels professional. Slick onboarding feels like a real company.

I think that's mostly wrong. The shape of products that age well — the ones I still use a year later, the ones I send other people — is almost the inverse. Narrower, more transparent, more honest, more identifiably human. PromptCraft tries to be that shape. So do the tools I keep coming back to.

If you're building one of those tools, I'd love to see it. The bar is high, but it's not because the world has too many AI tools. It's because most of them fail one of the five tests above, and the ones that pass are worth their weight in gold.