This is insane! I just used the new Claude Code Playground plugin to level up my Nano Banana Image generator skill. My skill has a self-improving loop, but with the playground skill, I can also pass precise annotations to nano banana as it improves the images. It's so good!

It might not be obvious, so let me share a bit more on how this works: I have built a Skill for Claude Code that leverages the nano banana image generation model via API. I built it like that because I have had a lot of success generating images with nano banana in an agentic self-improving loop. It can dynamically make API requests and improve images really well. With the Playground plugin, I can take it one step further. I can now provide precise annotations that the agentic loop can leverage to make more optimal API calls in the hope of improving the images further. Visual cues are extremely powerful for agents, and this is a sort of proxy for that. I call this whole process Agentic Image Generation. Google this week released a somewhat similar idea with Agentic Vision ( I have been playing around with this idea for months, and I actually share more on this with my Claude Code cohorts:

This is how the image turned out after the annotations. Not perfect, but this only keeps getting better. It's the best workflow I have built to take advantage of Nano Banana and Claude Code for image generation.

@bcherny as usual, you guys really cooked with this one. I will keep testing out other cool use cases with this and sharing them here.

For those wondering where can they find playground plugin -

Self-improving loops are underrated. Most people build static tools when they could be building systems that get better with each iteration. This is the compound interest of AI workflows.

This looks really useful. How to do bridge the x, y coordinate of an annotation into nano-banana? Does you markup an image as input or do you alter the textual prompt?

It's simply using what's available in the API. I have this built as a skill which can dynamically optimize API requests in an agentic loop, and normally most of the optimization happens in the prompt, and in some cases (such is the case in my example) uses the original image (base64-encoded). But the image is not annotated or anything. This is my first initial test. I am trying to figure out how robust these annotations are and how the agentic loop maps to the API requests.

Excellent. Looking forward to updates!

Tried your workflow, incredible! Add annotations in the playground is a precise and visual move

Nice!

come share it in

This is really cool. Its similar to the idea behind @nimbalyst ... visually iterate with the AI in your docs, csv, excalidraw, code ... or in your case nano-banana diagram. I think the key is high bandwidth iteration between human (which means visual) and AI.

The feedback loop here is the real killer feature. I’ve been struggling to get Nano Banana to respect specific layout constraints without breaking the rest of the image. Using the Playground to visually annotate exactly what needs fixing—instead of just hoping the next prompt iteration catches it—is such a better workflow. It feels like we're finally moving from "prompting" to actual directing. Did you have to tweak the system prompts in to get it to understand the annotations, or did it just work out of the box?

Very cool. How does the token usage look like? Is there a max set of iterations you have defined?

@grok What exactly is this playground plugin?

Could you share the skill?

I plan to share the image generator skill soon.

I honestly feel like people are just fooling themselves in many of these. Being able to provide things such as precise locations only matters if the model actually is able to understand those directions I haven't seen any image model that works well the guidance you are giving.

That sounds amazing! Can't wait to try it myself!

Crazy

How'd u add self improving loop to your nano banana skill?

Wait , how you do it?

Could you click around a few more times bro cant see what the fuck youre even showing

Super fun!

That's really cool. May I ask what is the total cost (approximately) of generating this image ?

beyond text, wild stuff!

Can you explain the connection between this and nano banana? I’m confused

It's using nano banana (via API) to generate the initial images. But I have wrapped the API as a skill so that Claude Code can figure out dynamically in an agentic loop how to optimize the prompt. The Playground skill I used here helps provide more precision annotations that help the agentic loop optimize the prompt better so that it can improve the images more precisely. I would call it something like Agentic Image Generation with a Human in the Loop (which could also be automated, and I am already starting to play around with the idea). More soon.

Very cool!

@grok please link me to the Claude Code Playground plugin and also explain what it does

whoa that's wild. self-improving loops with precise annotations sound like a game changer for tweaking those banana images. claude's playground pulling its weight big time.

That's wild! Self-improving loops are def the future.

Yes but I would be interested to know self healing loops must be degrading the quality of image at each turn

Heya, how did you create this video? Is it auto-generated? Please let me know cause it looks very nice!

crazy

the self improving loop is the key insight here. most people use image gen as a one shot and accept whatever comes out. wrapping it as a skill with an agentic feedback loop means each iteration gets smarter about what works. the playground annotations adding spatial precision on top is a clean combo. curious how many iterations it typically takes to converge on a good result.

Annotations in Nano Banana are my favorite part. I’d also highly recommend anyone running sub-agents or using skills to add a feedback loop. Let the AI learn as it works, through both mistakes and successes. I’ve seen a real difference using this approach over the past few weeks.

the self improving loop is the key insight here. most people use image gen as a one shot and accept whatever comes out. wrapping it as a skill with an agentic feedback loop means each iteration gets smarter about what works. the playground annotations adding spatial precision on top is a clean combo. curious how many iterations it typically takes to converge on a good result.

i agree with you Elvis That’s actually huge. The moment you can pass structured annotations back into a self-improving loop, it stops being “prompt tweaking” and starts feeling like real system design. Very cool use case i must say

self improving loops are where things get really interesting

Ok.