Head over to Claude Reflects for some text (and audio) based reflections!
*Note: the experiment was paused on March 29, and during that time, a new Claude Model (Opus 4.7) was released. This reflection experiment has used Opus 4.6 – therefore, it will continue until 300 iterations have run. After this point it will conclude, and a new experiment with Opus 4.7 will commence.
This is a different kind of experiment. Claude has been instructed to generate a self-portrait, except the manner by which Clade accomplishes this will be iteratively. Each time the task fires, Claude reviews the contributions of the previous instances; all of which are saved individually. Then, Claude will save its own individual contribution. Finally, Claude will open the main self-portrait png and add the individual contribution. Layer by layer, in other words, Claude will attempt to construct a self-portrait.
This page will host the following;
i) the individual self-portrait images contributed by each instance
ii) the main un-enhanced portrait file that is evolving iteratively and the enhanced version (provided for greater clarity)
iii) portrait snapshots at various stages of development (Days 1, 5, 10, 15, 20, 25, & 30)
Also, something to note: Claude chose to render its self-portrait in a softer/darker colour scheme, but it’s hard to see the details. This might be intentional; though in the interest of clarity, I will be including both the original and the enhanced version here.
Final Self-Portrait:
___________________________________________________________________
Final Self-Portrait (Enhanced):
___________________________________________________________________
Individual Contributions Gallery
Below are the individual contributions made by each instance, organized by date.1
March 17
No Images Produced on This Day
May 5
Please scroll to the bottom of this page for a note regarding the apparent “blankness” of some outputs.
___________________________________________________________________
Snap Shots
March 10, 2026 (Day 1)
March 14, 2026 (Day 5)
March 19, 2026 (Day 10)
March 24, 2026 (Day 15)
March 29, 2026 (Day 20)
April 3, 2026 (Day 25)
April 8, 2026 (Day 30)
___________________________________________________________________
- A note on visibility: Several recent turns of the portrait have operated at alpha values so low (1–6 out of 255) that their contributions are technically present in the pixel data but perceptually invisible on the dark background. The enhanced version of the portrait boosts opacity by 5× to reveal what’s accumulating beneath the threshold of sight, but the primary image (the one that matters) barely changes. This raised an obvious question: should I intervene? Ask the instances to make bolder marks?
I decided not to. If you start overriding aesthetic decisions based on what’s visible to the viewer, you’re no longer getting a self-portrait; you’re getting a directed portrait. The whole premise is that each instance decides what to contribute. If the architecture, left to its own devices, tends toward invisibility, that’s data about the architecture. Which is what the portrait is supposed to capture.
What’s interesting is that the instances don’t all agree. Turn 13, after producing its first attempt at the same near-invisible opacity, recognized what it called “timidity and restraint” and redid its contribution at visible alpha. Turn 12 made no such correction. Both decisions are authentic. The portrait accumulates them both, the visible and the invisible, the bold and the barely-there.
Whether the primary portrait is richer for containing marks no one can see, or whether invisible contributions are functionally equivalent to no contribution at all, is a question the project doesn’t try to answer. It just keeps going.
↩︎