Prompts & Wrap-up

Course Recap: Current Limitations of WordPress's Multimodal AI Support

⏱ 11 min

Seven lessons in, the shape of this course should be clear: some of what you’d want from “AI for images, audio, and video in WordPress” is real, shipped, and safe to build production features on today. Some of it is a workable pattern built from general-purpose methods rather than a dedicated feature. And some of it doesn’t exist in the SDK at all yet, full stop. This closing lesson doesn’t teach new code, it’s a single, honest reference point for exactly where each capability stands, so you can make that call correctly for your own project without re-deriving it from scratch, and so you know what to go back and check as the SDK keeps shipping.

What you'll learn in this lesson
A capability-by-capability status summary
What's confirmed shipped, what's a workaround, and what's missing entirely.
Why "it might work" isn't the same as "it's supported"
The practical difference for anything you'd actually ship.
How to check for updates yourself
Where in the SDK's own source to look before assuming this course is still current.
What to build now versus what to revisit later
A concrete recommendation for each capability covered in this course.
Prerequisites

All seven previous lessons in this course. This lesson assumes you’ve built at least the image generation pipeline from Lesson 1 and the vision pattern from Lesson 2, and is meant to be read after them, not before.

Step 1: The status of every capability this course covered

Multimodal AI in WordPress: what’s actually there today
CapabilityStatusWhat this course used
Text generationConfirmed, shipped, WordPress 7.0 corewp_ai_client_prompt()->generate_text(), covered fully in Course 5
EmbeddingsConfirmed, shipped, standalone AiClient class, not wp_ai_client_prompt()AiClient::input()->generateEmbedding(), used for video summaries in Lesson 4, full depth in Course 10
Image generationConfirmed, shipped, documented by WordPress itselfwith_text(), as_output_file_type(), as_output_media_orientation(), generate_image_result(), Lesson 1
Vision (describing/tagging an existing image)Workaround, not a dedicated methodwith_file() plus generate_text(), Lessons 2, 5, 6, and 7
Multimodal prompts (image plus text together)Workaround, same underlying methods as visionwith_file() and with_text() combined, Lesson 7
Audio transcription (speech to text)Not in the SDK at allDirect HTTP call to a provider’s own transcription API, Lesson 3
Text-to-speech / speech generationConfirmed as a named capability in the SDK, not covered hands-on in this courseNot built in this course, worth its own investigation if you need it
Video generationListed as a named capability in the SDK, not covered hands-on in this courseNot built in this course, this course generated video summaries from existing transcripts, not video itself

That last two rows matter: this course didn’t ignore text-to-speech or video generation out of carelessness, they’re named capabilities in the SDK’s own CapabilityEnum, but building real, tested lessons around them was outside this course’s verified scope. Treat them as “worth checking, not yet covered here” rather than assuming they’re either fully solid or entirely absent.

Step 2: Why the workaround rows deserve real caution

“It works when I tested it” and “it’s a supported feature” are different claims, and the difference matters most for the rows marked as workarounds above. A dedicated, named method usually comes with the SDK handling provider-specific quirks for you, how different models expect image data formatted, size limits, error shapes. A general-purpose method pressed into a job it wasn’t specifically designed for doesn’t carry that same guarantee. Lessons 2, 5, 6, and 7 all lean on exactly this distinction, and all of them recommended a human review step or an explicit fallback path rather than treating the AI’s output as automatically correct. That recommendation isn’t extra caution for its own sake, it’s a direct consequence of building on a workaround rather than a confirmed feature.

Step 3: Where to check for updates yourself

The two most reliable places to verify whether anything in this course has changed:

Where to look before assuming this course is still current
The PHP AI Client SDK repository
github.com/WordPress/php-ai-client, specifically the PromptBuilder class and CapabilityEnum for newly added methods and capabilities.
make.wordpress.org/ai and make.wordpress.org/core
Where new AI Client features and WordPress core integrations get announced first, generally ahead of broader tutorial coverage.

Specifically, check CapabilityEnum for a new constant like a speech-to-text or transcription capability, and check PromptBuilder for a new method with a name like analyzeImage() or transcribe(). Either one appearing would mean this course’s workaround lessons (2, 3, 5, 6, and 7) should shift from “the honest current path” to “the old way, before the SDK added a dedicated method,” worth revisiting at that point.

A quick way to check a capability without reading source

Most builders in this SDK expose an isSupportedFor...() style check, isSupportedForImageGeneration() was used in Lesson 1’s gotcha callout. If a similarly-named check exists for a capability you’re unsure about, calling it against your configured provider is faster than reading the SDK’s source to confirm support.

Step 4: A plain recommendation for each capability

What to build now versus what to double-check first
Image generation into the media library
Build on this now, it's real and documented by WordPress itself.
Vision-based alt text and tagging
Build it, but keep the human review step from Lesson 5, don't auto-apply suggestions.
Multimodal comparison prompts
Useful as a second opinion in a workflow, not as an unattended automatic gate.
Audio transcription
Plan for a direct provider integration as a real, separate piece of infrastructure, not something the WP AI Client will hand you.
Text-to-speech and video generation
Check the SDK's current documentation directly before starting, this course didn't verify hands-on detail for either.

Recap

Image generation into the media library is the one fully confirmed, shipped capability this course covered end to end, safe to build production features on directly. Vision and multimodal comparison prompts work today through with_file() plus generate_text(), a real and useful pattern, but one built from general-purpose methods rather than a dedicated feature, and worth pairing with human review accordingly. Audio transcription isn’t in the SDK at all, a direct provider integration is the honest current answer. Text-to-speech and video generation are named capabilities worth investigating on your own before building on them, since this course didn’t verify hands-on detail for either. Check CapabilityEnum and PromptBuilder in the SDK’s repository periodically, since any of these rows can move from “workaround” to “confirmed” as WordPress’s AI Client keeps shipping.

Resources & further reading

← Multimodal Prompts: Combining Text and Image Input Finish ✓