Which AI for Which Job: A Working Shortlist
Every model is best at something and mediocre at something else. This is which one I open for which task, based on using all of them inside real client work rather than on benchmarks.

The honest answer to "which AI is best" is that the question is wrong. I keep four or five open in a normal week and the choice depends entirely on the job in front of me. Asking which one is best is like asking which tool in a workshop is best.
So here are the actual jobs, and what I reach for in each. I do QA work, WordPress and PHP development, and UX/UI design, so this list is skewed toward that. It is not a benchmark table. It is what I open.
Writing code from scratch
Claude, in an editor rather than a chat window. The difference that matters here is not raw model quality, it is context: a tool that can read the whole project produces code that fits the project. A chat window that sees one pasted file will confidently reinvent a helper you already have.
If the task is a self-contained script with no surrounding codebase, any of them will do and the difference is not worth thinking about.
Reviewing code somebody else wrote
This is where I get the most value and it is the use nobody markets. Paste an unfamiliar function, ask what it does, ask what breaks it. For WordPress specifically, ask directly whether the nonce and capability checks are there, because the model will not raise it unprompted.
Claude and ChatGPT are both good here. What matters more than the choice is the question: "explain this" gets you a summary, "what input breaks this" gets you something useful.
Test case design
Any of the frontier models, and none of them alone. I write my own risk list first, then ask for cases from the same requirement, then merge. Reversing that order produces thorough-looking coverage of the wrong things, which I go into in AI for manual testers.
Fast UI prototypes
Lovable or Bolt.new. These generate a working front end from a description in minutes, which is genuinely useful for showing a client an idea instead of describing it. The code underneath is throwaway. Treat the output as a clickable wireframe, not as the start of your production app, and it earns its place.
Visual design work
Almost none of them, honestly. AI is useful for generating variations on something that already exists and for the tedious parts of production work: resizing, alt text, first-draft copy for a component. Deciding what the interface should be is still the job, and that decision is where design actually lives.
Figma's built-in AI features are more useful than a general chat model here, because they operate on the file rather than on a description of the file.
Research and reading around a problem
A model with live search. The value is not the answer, it is the set of sources it surfaces that I then read myself. Any model answering a technical question from training data alone will be confidently out of date on anything that moved in the last year, which in this field is most things.
Client-facing copy
First draft only, then rewrite. AI copy has a recognisable texture: heavy on em dashes, fond of words like seamless and comprehensive, every paragraph the same length. Clients may not name it but they feel it. If you publish it unedited across a site you also run into Google's position on scaled content.
Explaining something to a non-technical client
This one surprised me. Asking a model to explain a technical constraint in plain language, for someone with no background, produces better analogies than I come up with under time pressure. I still rewrite it in my own voice, but the framing is often better than mine.
Deciding anything
None of them. Not whether a bug is a defect, not whether a design works, not whether a client's requirement makes sense. Models answer decisively regardless of whether they have grounds to, and a confident wrong answer at a decision point costs more than no answer at all.
The pattern underneath
Every use above where AI helps has the same shape: it does something mechanical, and I check it. Every use where it hurts has the other shape: it makes a judgement, and I accept it.
That is a more useful filter than any comparison table, and it survives the models changing every three months.
Related
The risks side: vibe coding risks. WordPress specifics: AI for WordPress. Design specifics: AI in UX/UI design.
Digər yazılar
AI can write a WordPress plugin in a minute. Whether you should put that plugin on a live site is the part nobody covers. From someone who reviews this code for a living.
Most slow WordPress sites are slow for two or three specific reasons. Here is how to find which ones, in the order that gives you the biggest drop in load time for the least work.
AI test keyslərini saniyələr içində yaza bilir. Həmin keyslərin nəyisə tapıb-tapmayacağı ayrı sualdır. Real QA işində bir illik istifadənin nəyi dəyişdiyi.
AI kodu saniyələr içində yazır. Onun işlədiyini kiminsə yenə yoxlaması lazımdır. Vibe coding-in əsl qiyməti və AI alətlərini onların səhvlərini dərc etmədən işlətmək.
On UX/UI trendi, amma sıralanma prinsipi başqadır: qurduğun məhsulu dəyişirmi, yoxsa sadəcə bu il dizayn bloglarının görünüşünü?
WordPress saytını qorumaq üçün işlək siyahı: giriş, fayllar, baza, hostinq və monitorinq. Nəzəri deyil, real saytlarda nəyin sındığından yazılıb.
AI ekranı düzə bilir, davranış məlumatını oxuyur, istifadə yoxlamaları aparır. Məhsulun nə olmalı olduğuna qərar verə bilmir. Sərhədin harada olduğuna praktik baxış.