Two things become obvious once you start writing WordPress code alongside an AI. The first is that the tedious-to-memorise material — hook names, function signatures, argument order — genuinely gets faster. The second is that there are stretches where no amount of re-prompting stops it returning code that has nothing to do with your project. The cause of the second is rarely the model. It is how much you handed over.
Ask without context and you get the generic answer
Ask “add an admin settings page” and what comes back tends to look like the example in the handbook. Function names carry no prefix, the option key is something like my_plugin_options, and the save handler uses the most widely published approach rather than yours. None of that is wrong. It simply does not fit any particular project, including yours.
The usual reactions are “try a bigger model” or “write a longer prompt”, and both miss the point. The model has never seen your repository. It does not know your prefix convention, the PHP version you have to support, or the helper sitting fifty lines above the insertion point. There is no way to make it guess unknown facts well — only ways to tell it.
Four things to send with the request
Conventions fit in a paragraph: the prefix for functions and option keys, the text domain, where new files go, how output is escaped. Those four lines alone make the returned identifiers look like the rest of the repository, and that similarity removes most of the review time.
Version constraints are the item whose absence is discovered latest. Code that runs locally throws a parse error on the supported PHP version, or calls something recently deprecated, or calls a function your WordPress release does not have yet. One line — “must run on PHP 7.4 and above” — filters out that entire class.
Surrounding code has the largest effect of the four. Paste the region you are editing and the model calls your existing helper instead of writing a second one. Withhold it and you get a duplicate, and duplicates cause trouble later in the specific way that only one of the two ever gets fixed.
How it will be checked turns the request into something verifiable. Write “passing means wp option get acme_settings shows the value after saving” and the output aims at that condition — and you already know what you are about to test.
The same request, two ways
Side by side, the difference is not subtle.
[thin request]
add an admin settings page
[request carrying context]
- plugin acme-core, function prefix acme_, text domain acme
- supports PHP 7.4+ (no enums, no readonly)
- settings live in one option array acme_settings, autoload=no
- save via admin_post + check_admin_referer + current_user_can( 'manage_options' )
- all output goes through esc_html / esc_attr
- goes in includes/class-settings.php (current contents attached)
- passing: after saving, wp option get acme_settings shows the valueEvery line of the second request is a condition you can verify in the resulting code. That is the test for good context: if, having read a line, you cannot answer “how would I know whether this was honoured?”, it is still a preference rather than a requirement.
You do not have to retype that list each time. Conventions and constraints are near-constant per project, so they belong in a document in the repository that gets read at the start of every session. That is how we work, and the structure of such a document deserves its own discussion.
What to check first in what comes back
Good context does not remove review. It does decide what you look at first.
Non-existent hook names deserve special mention because they fail silently. add_action() accepts any string, so a callback attached to a mistyped hook never runs and never complains. Whenever you attach a new hook, confirm once that it actually fires.
The short version: what you scale up in AI collaboration is not prompt length but the number of facts you hand over. Conventions, versions, surrounding code, and how it will be checked — four lines that visibly change the output, and that you write once and keep using.
Where and how to attach hooks is covered in the Using hooks and filters properly series, and other ways of folding AI into a workflow live in the AI in practice archive. If the thing that needs work is a running site rather than a file, our optimization program handles upgrades, performance and security as one job.