Compatibility List

The cards I can and can't support

By far, the most challenging part of this project is handling the near infinite variety of Anki note types. People use Anki in wild and creative ways. I have battle-tested the system against thousands of note types, but there are still limitations outside my control and edge cases I have not yet encountered.

While I support custom note types and cloze deletions, I have reason to believe that “basic” note types are the most optimal (image occlusion is fine too I guess). I think all of the other features and crazy note types that people create are a distraction and a waste of time. Just lock in and KISS KISS.

If you aren't confident I can render your card properly, I recommend starting with a small audit in Sample Mode to test for audit quality.

Scenario	Status	Notes
File Formats
.apkg	Supported
Images
.png / .jpg / .jpeg / .webp	Supported	Supported. Per-image limit is 7 MB and 3072px on any side. Larger images are automatically resized before being sent to the AI auditor. Images smaller than 50px are upscaled to 100px to improve readability. Cards are capped at 20 images. If a card has more than 20 images, the extras are dropped before sending to the AI auditor. This limit exists as a safeguard against excessive processing costs.
.gif	Supported	GIFs are rendered natively by the browser just like they would appear in Anki. Animated GIFs are re-rendered into short video clips so the LLM sees the animation, not just a frozen frame. Static GIFs are treated as regular images.
.svg	Supported	SVGs render natively in the headless browser, and the AI auditor evaluates them through the card screenshot. Complex SVGs with embedded scripts may not render perfectly.
Video & Audio
.mp4 / .mov / .avi / .webm / .mkv / .flv	Supported	Video files are transcoded to H.264 MP4 and sent to the AI auditor as multimodal input. Cards are capped at 3 videos; extras are dropped before auditing. Most flashcard videos are short clips — longer videos work fine but are best kept under 30 seconds. The visual render of the card will show a placeholder where the video would appear.
.mp3 / .wav / .ogg / .m4a / .flac / .aac	Supported	Audio files are transcoded to MP3 and sent to the AI auditor as multimodal input. Cards are capped at 5 audio files; extras are dropped before auditing. Most flashcard audio is a short pronunciation clip or sentence — longer clips work fine but are best kept under 1 minute. The AI can identify spoken language content (e.g., pronunciation guides, vocabulary audio).
Note Types & Features
Basic note type	Supported	The GOAT note. Use basic as much as possible. Focus on studying not inventing new note types.
Cloze deletions	Supported	Supported. Each cloze card is rendered and audited individually. The AI auditor sees each deletion in isolation, which is the intended Anki behavior.
Custom note types	Partial	If you have custom note types with various fields and templates, this is most likely supported unless you are doing super weird stuff with JavaScript. The renderer injects the `#qa` and `#qa_box` containers that most community templates expect.
Native image occlusion	Supported	Anki's built-in native image occlusion is supported for rendering and auditing.
Image Occlusion Enhanced	Supported	The Image Occlusion Enhanced add-on is supported for rendering and auditing.
Math & Typesetting
MathJax / LaTeX	Supported	Both are supported. MathJax is bundled directly into the render worker, no internet required. Anki's legacy delimiters (`[latex]`, `[$]`, `[$$]`, `[chem]`) are automatically converted before rendering.
Custom fonts	Supported	Fonts included in your APKG media folder are automatically detected and injected into the render. If a card uses a custom font, it will look the same as it does in Anki.
Rendering Behavior
Custom CSS	Partial	Mostly supported. Custom styling is passed through with minimal transformation. The vast majority of CSS works correctly. Edge cases involving `position: fixed` or viewport-relative units may produce unexpected layouts in the screenshot.
Custom JavaScript	Partial	Custom template scripts run in the headless Chrome sandbox. Each script is wrapped in a `try/catch`so a broken script won't kill the rest of the render. Cards that rely on asynchronous JS (e.g., `setTimeout`-based animations) may not render as expected since the screenshot is taken when the page loads, not when JS finishes.
External network calls	Not supported	If your card has JavaScript that fetches an image or an online resource dynamically, this is not supported. The visual render worker does not have any kind of internet access for security reasons. This is generally a bad idea though, don't do this.
AI Limitations
Hearing sounds (guitar chords, bird calls, songs)	Partial	Audio is always sent to the AI auditor. Speech-based audio (pronunciation, spoken words) works well. Non-speech sounds like guitar chords, bird calls, or instrumental music may not be recognized accurately — this is a current AI limitation that is improving over time.
Very recent events	Not supported	LLMs have a knowledge cutoff date. If you are asking for knowledge that can only be known post the cutoff date, the AI auditor will just assume the content of the card is correct. Gemini's cutoff can be found in Google's model documentation. Parametric knowledge cutoffs will likely shrink over time.
Inappropriate content / PII	Not supported	People should be free to use Anki to memorize whatever they want, but the system will flag these automatically, and not display them in the audit results. No refund is issued for flagged cards.

Scenario

Status

Notes

File Formats

.apkg

Supported

Images

.png / .jpg / .jpeg / .webp

Supported

Supported. Per-image limit is 7 MB and 3072px on any side. Larger images are automatically resized before being sent to the AI auditor. Images smaller than 50px are upscaled to 100px to improve readability. Cards are capped at 20 images. If a card has more than 20 images, the extras are dropped before sending to the AI auditor. This limit exists as a safeguard against excessive processing costs.

.gif

Supported

GIFs are rendered natively by the browser just like they would appear in Anki. Animated GIFs are re-rendered into short video clips so the LLM sees the animation, not just a frozen frame. Static GIFs are treated as regular images.

.svg

Supported

SVGs render natively in the headless browser, and the AI auditor evaluates them through the card screenshot. Complex SVGs with embedded scripts may not render perfectly.

Video & Audio

.mp4 / .mov / .avi / .webm / .mkv / .flv

Supported

Video files are transcoded to H.264 MP4 and sent to the AI auditor as multimodal input. Cards are capped at 3 videos; extras are dropped before auditing. Most flashcard videos are short clips — longer videos work fine but are best kept under 30 seconds. The visual render of the card will show a placeholder where the video would appear.

.mp3 / .wav / .ogg / .m4a / .flac / .aac

Supported

Audio files are transcoded to MP3 and sent to the AI auditor as multimodal input. Cards are capped at 5 audio files; extras are dropped before auditing. Most flashcard audio is a short pronunciation clip or sentence — longer clips work fine but are best kept under 1 minute. The AI can identify spoken language content (e.g., pronunciation guides, vocabulary audio).

Note Types & Features

Basic note type

Supported

The GOAT note. Use basic as much as possible. Focus on studying not inventing new note types.

Cloze deletions

Supported

Supported. Each cloze card is rendered and audited individually. The AI auditor sees each deletion in isolation, which is the intended Anki behavior.

Custom note types

Partial

If you have custom note types with various fields and templates, this is most likely supported unless you are doing super weird stuff with JavaScript. The renderer injects the #qa and #qa_box containers that most community templates expect.

Native image occlusion

Supported

Anki's built-in native image occlusion is supported for rendering and auditing.

Image Occlusion Enhanced

Supported

The Image Occlusion Enhanced add-on is supported for rendering and auditing.

Math & Typesetting

MathJax / LaTeX

Supported

Both are supported. MathJax is bundled directly into the render worker, no internet required. Anki's legacy delimiters ([latex], [$], [$$], [chem]) are automatically converted before rendering.

Custom fonts

Supported

Fonts included in your APKG media folder are automatically detected and injected into the render. If a card uses a custom font, it will look the same as it does in Anki.

Rendering Behavior

Custom CSS

Partial

Mostly supported. Custom styling is passed through with minimal transformation. The vast majority of CSS works correctly. Edge cases involving position: fixed or viewport-relative units may produce unexpected layouts in the screenshot.

Custom JavaScript

Partial

Custom template scripts run in the headless Chrome sandbox. Each script is wrapped in a try/catchso a broken script won't kill the rest of the render. Cards that rely on asynchronous JS (e.g., setTimeout-based animations) may not render as expected since the screenshot is taken when the page loads, not when JS finishes.

External network calls

Not supported

If your card has JavaScript that fetches an image or an online resource dynamically, this is not supported. The visual render worker does not have any kind of internet access for security reasons. This is generally a bad idea though, don't do this.

AI Limitations

Hearing sounds (guitar chords, bird calls, songs)

Partial

Audio is always sent to the AI auditor. Speech-based audio (pronunciation, spoken words) works well. Non-speech sounds like guitar chords, bird calls, or instrumental music may not be recognized accurately — this is a current AI limitation that is improving over time.

Very recent events

Not supported

LLMs have a knowledge cutoff date. If you are asking for knowledge that can only be known post the cutoff date, the AI auditor will just assume the content of the card is correct. Gemini's cutoff can be found in Google's model documentation. Parametric knowledge cutoffs will likely shrink over time.

Inappropriate content / PII

Not supported

People should be free to use Anki to memorize whatever they want, but the system will flag these automatically, and not display them in the audit results. No refund is issued for flagged cards.