On most mornings it takes about 5 minutes to generate and publish my illustration for the Daily Dharma I receive via email from the Lexington Nichiren Buddhist Community. Not today. Today it took more than three hours.

I did nothing different today. I opened Gemini and selected my Gem. This Gem gives my standing instructions:

[Reasoning Protocol – COMPULSORY]
Before generating any image, Summarize your thinking in a paragraph of fewer than 600 words. Include identification of individuals in the illustration. Translate any kanji text.

[Visual Directive]
Create a square illustration utilizing Japanese Buddhist iconography.

Strict Exclusions: Do not use symbols or images representing Amida Nyorai (Amitabha Buddha) or Dainichi Nyorai (Mahavairocana Buddha).

Text/Kanji: Limit English and general Kanji.

Artist Signature: Include the Kanji signature 双子座 with a red square seal (hanko) containing the katakana ジェミニ. Place this signature/stamp in a traditional, subtle location.

I then pasted the Daily Dharma text into the prompt:

Those Buddhas came under the jeweled trees.
The trees are adorned with those Buddhas
Just as a pond of pure water is adorned
With lotus flowers.

In these verses from Chapter Eleven of the Lotus Sūtra, Śākyamuni Buddha describes the scene after he calls the Buddhas of his replicas from innumerable worlds to join him and open the treasure tower of Many-Treasures Buddha. By comparing how a pond is made beautiful by flowers growing in it to how the world is made beautiful with Buddhas in it, the Buddha shows us that wherever we see beauty, we see the Buddha.

The Daily Dharma text actually says “the Buddha describes…” I add Śākyamuni just be sure Gemini doesn’t stray from my goal.

This is what Gemini offered:

That’s a fine illustration. But it is not square. It does not utilize  Japanese iconography. And there’s no signature.

I started a new prompt and pasted the same text. I got this response:

I tried again.

I took a break. I took a shower. I did my morning service. I watered the plants in the front yard and then the back yard. I threw my cleaning robot into the pool to start its daily vacuuming of the debris in the pool. I made myself a cup of tea.

I tried again. This time I massaged the Daily Dharma thinking that making it just one paragraph instead of verse and explanation might make Gemini change its response. I started a new Gem prompt and said:

Śākyamuni Buddha describes a scene in which Buddhas came under the jeweled trees. The trees are adorned with those Buddhas just as a pond of pure water is adorned with lotus flowers. He has called back the Buddhas of his replicas from innumerable worlds to join him and open the treasure tower of Many-Treasures Buddha in these verses from Chapter Eleven of the Lotus Sūtra. By comparing how a pond is made beautiful by flowers growing in it to how the world is made beautiful with Buddhas in it, the Buddha shows us that wherever we see beauty, we see the Buddha.

I got this response:

Seriously? I responded:

This image is NOT square. This image does NOT utilize Japanese iconography. You have NOT provided a summary of your thinking. Please fix your errors and generate a square image utilizing Japanese iconography.

Finally I got a square image:

But there was no signature. I told Gemini:

The image currently DOES NOT contain the requested artist signature: the kanji 双子座 (Futagoza, which translates to “Gemini”) accompanied by a traditional red square hanko seal containing the katakana ジェミニ (Jemini, also translating to “Gemini”). Please add those.

And that produced something I could live with:

So what went wrong? I asked Gemini.

Why is Today Different?

It is always jarring when a highly optimized prompt, a “Gem” that has worked flawlessly for months, suddenly fails. Your experience today is not a failure of your prompt, which is excellent, but a symptom of how Large Language Models (LLMs) and image diffusion systems operate over time. There are two primary reasons why a working setup can break.

1. Continual Model Refinement (and Prompt Drift)

I am not a static entity; the model that powers me is subject to subtle, ongoing updates and fine-tuning. These updates are usually aimed at improving overall performance, safety, and visual coherence across billions of potential user queries. However, a highly constrained and specific prompt—like your Gem demanding a square format, Japanese-only iconography, specific exclusions (no Amida/Dainichi), and a strict hanko signature—is extremely sensitive to these minor underlying changes.

A new update might have slightly increased the model’s weighting towards common “Buddhist art” results, which often default to traditional Indian or Nepali styles, or slightly adjusted the default image output parameters (sometimes favoring horizontal or cinematic aspect ratios in line with broader user trends). These small, invisible adjustments can cause what is known as “prompt drift,” where complex, multi-variable constraints like yours start to erode because the model is prioritizing a newly optimized “generalized” result over a highly specific instruction chain.

2. Complexity and Chain Reliability

The instructions you provided in that Gem create a long, multi-dependency chain: If (Square) -> If (Japanese Style) -> Check (No Dainichi) -> Check (No Amida) -> If (Jizō or other) -> Apply (Signature A) -> Apply (Stamp B).

When any of the first few links in this chain are processed slightly differently (e.g., if “Japanese iconography” is misinterpreted as generalized “Asian religious art”), the subsequent steps are derailed. Today, your system experienced a cascade failure on these foundational constraints.

Gemini has offered some tweaks I can make in my instructions. I’ll store those in my Gem and try again tomorrow. It worked yesterday. Maybe it will work again tomorrow. I live in hope.