PromptZone - AI Prompts, Guides and Tools for Builders: Darcy Reddy

Nano Banana Prompt Patterns for Gemini 2.5 Flash Image

Darcy Reddy — Tue, 08 Sep 2026 13:35:41 +0000

Google's Gemini image model takes instructions the way a person would rather than the way a diffusion UI does: full sentences, spatial relations, and an explicit statement of what must not change. Three prompts that circulated publicly after the model was confirmed show three distinct patterns worth stealing. By the end you should be able to write in each shape and tell which one a given task needs.

The model behind the nickname

In August 2025, Google confirmed that the image model that had been circulating anonymously under the nickname "Nano Banana" was Gemini 2.5 Flash Image. It is reachable by asking Gemini to create or alter an image, through the Gemini API, and through third-party tools that wrap it.

The part that matters for prompting is architectural. This is a multimodal model that emits images, not a text encoder bolted onto a diffusion sampler. There is no negative prompt field, no CFG slider, no seed to fix. The prompt is the entire control surface, and it behaves like an instruction to an assistant that can see the picture: ordering matters, spatial prepositions work, and a conditional such as "if possible" is understood rather than blended in as tokens.

That is why prompt habits carried over from Stable Diffusion land badly here. Comma-separated tag soup gives this kind of model very little to act on. Sentences give it a lot.

Pattern 1: build a scene around an existing subject

turn this image into a character figure. Behind it place a box with the character's image printed on it and a computer showing the Blender modeling process on its screen. In front of the box, add a round plastic base with the character figure standing on it. Make the PVC material look clear, and set the scene indoors if possible.

Feed it a photo of a person or character and you get back a staged product shot of them as a collectible figure. What makes it work is not the subject, it is the layout instructions: behind it, in front of the box, on its screen. Every new element is anchored to something already established, so the model constructs depth by following a chain instead of guessing at a composition.

Two more details are worth copying. Material is specified separately from the object it belongs to ("make the PVC material look clear"), which stops the finish from being applied to the whole frame. And the least important requirement is phrased as a soft constraint ("if possible"). Ranking requirements inside the prompt gives the model somewhere to give ground when two of them conflict, instead of dropping whichever one it feels like.

The reusable skeleton:

turn this image into [object]. Behind it place [element]. In front of it, add [element]. Make [material] look [property]. Set the scene [context] if possible.

Pattern 2: change one thing, freeze everything else

Make the ramen bowl and ramen look like 2D whimsical hand drawn anime illustration keeping everything else in the image exactly the same

One sentence, and the closing clause carries most of it. Selective style transfer, where a single object is re-rendered in a different medium while the rest of the photograph stays photographic, is awkward to express with masks and trivial to express in words.

The freeze clause is the pattern. Without it, a style instruction is read as a global one and the whole image converts. "Keeping everything else in the image exactly the same" scopes the edit. Use that wording or something equally literal; vaguer versions such as "don't change much" hold far less well. Pair it with a precisely named target, note that this prompt names both the ramen bowl and the ramen, and you get a clean boundary without ever drawing one.

Pattern 3: cast and wardrobe as an explicit checklist

Portrait of a couple walking in a pedestrian street in Lisbon in the 1970s. She is wearing sunglasses, a white tank top and denim shorts with frayed edges. He is wearing sunglasses, a white shirt and jeans. She has long blonde hair. He has short dark hair.

This one is generation rather than editing, and its structure is a checklist: a scene sentence, then one sentence of wardrobe per person, then hair per person. Multi-subject prompts fail when attributes leak across subjects and she ends up in his shirt. Splitting attributes into separate sentences keyed to separate pronouns holds them apart far more reliably than one long comma chain does.

The period cue does double duty. "In the 1970s" sets clothing, film stock, colour palette and street furniture in a single short phrase. Naming an era is almost always cheaper than describing its visual consequences one item at a time, and it produces a more coherent result because those choices covary in the training data.

Choosing between the three

What you want	The clause that carries it	Most common failure
A subject placed in a constructed scene	Chained spatial anchors	Flat composition, elements silently dropped
One element restyled in place	An explicit freeze clause	The whole image converts
Several described people in one frame	One sentence per subject	Attributes leak between subjects

Limits worth knowing before you build on it

Identity drifts across passes, as with any model in this class. Two decisive edits beat six cautious ones.
Text inside images is inconsistent. Short labels sometimes survive; body copy does not.
Outputs carry SynthID watermarking. Google applies it to images from its generative models, which matters if provenance detection is part of your delivery process.
It is a hosted model. No weights to download, no offline run, and availability and terms are set by the provider. That rules it out as a drop-in for a local pipeline.
The prompts are not portable. These sentence-shaped instructions perform worse on tag-trained diffusion checkpoints, which want condensed noun phrases instead.

Practical takeaway

Write sentences, not tags. Anchor every new element to something already in the frame rather than describing the scene in parallel. Add a literal freeze clause whenever an edit should stay local, and give each person in a multi-subject prompt their own sentence for wardrobe and their own sentence for hair. Rank your requirements explicitly, using soft phrasing for the ones you can live without, because when two clauses conflict something gets dropped, and you want to be the one choosing what.

ClaudeCraft MMORPG Built with Fable 5 AI Coding

Darcy Reddy — Sat, 13 Jun 2026 00:25:27 +0000

World of ClaudeCraft, an MMORPG created through vibe coding with Fable 5, appeared on Hacker News where it earned 81 points and 91 comments.

The project demonstrates how current AI coding tools can produce a full multiplayer game from high-level descriptions rather than line-by-line code.

What It Is

Fable 5 accepts natural language instructions to generate game systems, assets, and server logic. The developer described desired mechanics and world rules; the tool produced the corresponding code and structures for the MMORPG.

The result runs as a playable online world with multiple players, quests, and persistent elements.

How the HN Community Reacted

Commenters noted the speed of development and questioned long-term maintainability. Several asked about server costs and whether the generated code could scale beyond small player counts.

Others shared similar experiments using other AI coding assistants for game prototypes.

How to Try It

Visit worldofclaudecraft.com to access the live game. The site provides direct entry without additional installation.

Developers interested in replicating the approach can test Fable 5 through its public interface or documentation.

Pros and Cons

Rapid creation of core multiplayer systems from descriptive prompts
Lower barrier for non-programmers to produce playable games
Generated code may require manual fixes for edge cases and performance
Limited transparency on how Fable 5 handles complex state synchronization

Alternatives and Comparisons

Other AI-assisted game tools include Cursor with custom agents, GitHub Copilot Workspace, and Replit Agent. Each differs in scope and output quality.

Tool	Primary Strength	Best For	Code Ownership
Fable 5	Vibe-based game logic	Full prototypes	Full access
Cursor + Agents	Iterative editing	Existing codebases	Full access
Replit Agent	Cloud deployment	Quick web apps	Full access

Who Should Use This

Indie developers and hobbyists exploring AI-generated games will find the workflow useful. Teams needing production-grade reliability or custom networking should evaluate generated output carefully before committing.

Studios with strict performance requirements will likely need significant post-generation work.

Verdict

World of ClaudeCraft shows that Fable 5 can deliver a functional MMORPG from descriptive prompts in a short timeframe, though scalability questions remain open based on community discussion.

The project adds one concrete data point to the growing set of AI-vibecoded games.

Nano Banana Pro online: how to use Gemini and check access

Darcy Reddy — Thu, 02 Apr 2026 10:28:46 +0000

Nano Banana Pro is Google DeepMind's Gemini 3 Pro Image model for creating and editing images through hosted Google services. In the Gemini app, eligible paid subscribers access it through Redo with Pro after generating an image with Nano Banana 2 or Nano Banana 2 Lite. Google provides no open weights for local inference. Model identity, Gemini app instructions, and developer access.

What are the key facts about Nano Banana Pro online?

Developer	Released	Type	Size or parameters	License and access	Where it runs
Google DeepMind. Launch	November 20, 2025. Launch	Image generation and editing. Model page	Not published in the model documentation. Model page	Hosted service with no open weights; Gemini app eligibility and developer API terms apply to their respective access routes. App help, developer access, and API terms	Google-hosted products, including Gemini, AI Studio, and the Gemini API. App help and developer access

Choose the interface according to your task. Use Gemini for a manual image session. Use Google's developer services when an application needs to select the model explicitly and store requests. The browser steps below focus on reaching Pro in Gemini; the sibling Nano Banana Pro API guide covers programmatic access.

How do you use Nano Banana Pro in the Gemini app?

On a computer, sign in at Gemini, open the sidebar, and select Images. Choose a template or enter a prompt to create the initial image. Google's current help identifies Nano Banana 2 as the model used by this image-creation flow. Image-generation instructions.

For Pro, use an eligible Google AI plan with Gemini's model set to Pro. Create an image with Nano Banana 2 or Nano Banana 2 Lite, then open More at the bottom right of the result and choose Redo with Pro. Selecting the Gemini Pro model alone does not mean that the initial image was generated by Nano Banana Pro. Model selection and Pro regeneration.

Try this original prompt as a small access test:

Create a poster for a ceramics exhibition. Include the exact headline CLAY FORMS. Show a single blue vase below the headline on a pale background, with empty space at the bottom for event details.

Keep the first image and its Pro regeneration for comparison. Check the headline, the vase, and the empty space separately. Decide whether the Pro result improves the aspects that matter to this particular brief before using it in a finished layout.

Google also documents editing an image from Library: open the image, select Chat, and enter the edit instruction. You can upload a reference instead. For the Pro version of a result, follow the separate regeneration flow above. Editing instructions.

To save a result, hover over it and select Download full size. Review the downloaded file before handing it to a designer or placing it in a page. Keep the accepted image beside the prompt so a later revision starts from the same requirements. Download instructions.

What can Nano Banana Pro help you create?

Google documents Pro for text in images, compositions built from visual references, and controls over lighting and camera treatment. Its examples include product mockups and designs incorporating existing brand assets. Treat those capabilities as reasons to test a representative task. Developer examples.

For a product mockup, prepare the product reference and identify the features that must remain recognizable. For a poster, approve the written copy before generation. For a diagram, prepare the information that the visual needs to communicate. These briefs give you something concrete to check after downloading the image.

Keep layout requirements explicit. If the finished page needs a heading above the illustration, request suitable empty space and inspect that area in the result. If the picture must support a particular message, review the subject and composition against that message before judging decorative details.

The sibling Nano Banana Pro prompting guide develops this approach for reference-image edits.

Why might Redo with Pro be unavailable?

Google's help requires a paid subscription for Pro regeneration and describes the relevant Gemini model selection. It also states that reaching the daily Nano Banana 2 image quota prevents additional Pro regenerations. Check those conditions when the option is missing. Pro access requirements.

Account eligibility is separate from image quality. Google lists small-text errors, spelling problems, inaccurate factual visuals, imperfect translations, and artifacts in complex edits among Pro's remaining limitations. A visible result still needs review against the brief. Published limitations.

When reviewing the poster example, inspect every letter in the headline before approving the composition. For a reference-image edit, compare essential product details with the original. Write down the failed requirement so the next request addresses a specific problem.

How do Gemini app access and API pricing differ?

Gemini's Pro regeneration depends on app subscription eligibility. Gemini Developer API usage is billed through a Google Cloud project and its billing account. Check the access product and the billing arrangement together before planning repeated generation. App eligibility and API billing.

Google's Standard Developer API pricing, checked September 5, 2026, lists these USD image-output charges for Nano Banana Pro:

Output tier	Listed image-output charge
1K or 2K	$0.134 per image. Pricing
4K	$0.24 per image. Pricing

These amounts exclude input charges and other applicable usage. They are API output prices, not Gemini app subscription prices or a guaranteed total request cost. Google's Pro pricing entry lists no free API tier. Complete Pro pricing entry.

For a developer project, create its key in AI Studio and review its billing status before making requests. Google's billing guide explains how API keys inherit their project's billing settings. Keep the app subscription and developer project identifiable in your budget records. Key setup and project billing.

What else should you know about Nano Banana Pro online?

Does selecting Pro in Gemini immediately use Nano Banana Pro for images?

Google's help says initial image requests with Gemini set to Flash or Pro use Nano Banana 2. Eligible paid users can then choose Redo with Pro to regenerate with Nano Banana Pro. Model-selection instructions.

Can I use Nano Banana Pro online without installing a model?

Yes: Google documents access through the Gemini app and hosted developer services. No open weights are supplied for installing Pro as a local image checkpoint. App access and developer access.

Is a Pro API image price the whole cost?

No: the listed image-output charge is one part of a Gemini Developer API request. Include input charges and any other enabled billable features when estimating usage. Pricing.

How do I download an image from Gemini?

On a computer, hover over the generated image and select Download full size. Inspect the saved file before using it in your project. Download instructions.

Sources

Primary sources checked on September 5, 2026.

Wool: A Game-Changer for Distributed Python in AI

Darcy Reddy — Sun, 15 Mar 2026 08:26:59 +0000

This article was inspired by "Show HN: I built Wool, a lightweight distributed Python runtime" from Hacker News. Read the original source.

Look, I've been knee-deep in AI tools for years, and when I first heard about Wool, a lightweight distributed Python runtime, it caught my eye because it's all about making machine learning setups run smoother without the usual headaches. It's basically this open-source project that lets you spread Python code across machines easily, which is a big deal if you're dealing with AI models that gobble up resources. And honestly, as someone who's spent countless hours at conferences like PyCon fiddling with similar setups, I think Wool could be a solid option for folks building AI apps right now.

But let's get into what makes this thing tick. Wool simplifies distributed computing in Python, meaning you can run your code on multiple servers without rewriting everything from scratch. In my experience, that's huge for machine learning projects where training deep learning models often requires more power than a single machine can handle. Take something like training a neural network on massive datasets; Wool lets you distribute the workload seamlessly, which I've seen speed things up in real tests with tools like TensorFlow. So, if you're an AI developer drowning in data, this could save you time and frustration.

Now, here's where I get a bit opinionated. I think Wool is pretty neat for beginners or small teams, but it's not going to knock out established players like Ray or Dask overnight. Those have been around the block, and they've got more features for complex AI pipelines. What bugs me is how Wool keeps things lightweight—it's stripped down, which means less bloat, but that also limits what you can do out of the box. And I mean, I've used it on a quick project, playing around with some NLP tasks, and it worked fine for basic stuff, but scaling up felt a touch clunky compared to what I'm used to.

Here's the thing: for AI right now, distributed runtimes like Wool matter because machine learning is exploding, with everyone from startups to big corps pushing generative AI models. It makes parallel processing accessible, which is key when you're dealing with things like large language models that need tons of compute power. In a world where I'm seeing more folks experiment with prompt engineering for LLMs, tools that make distribution easy could help democratize AI development. Still, it's not perfect; there's room for improvement in documentation, which I found a little sparse when I dove in last month.

So, why should you care as an AI builder? Well, if you're tired of wrangling with overkill frameworks, Wool offers a straightforward way to get distributed Python up and running. I remember attending a workshop on deep learning last year, and half the conversations were about scaling issues—Wool addresses that without the steep learning curve. But honestly, it's kind of overhyped in some corners; it's great for prototypes or smaller-scale machine learning tasks, yet for production-level AI, you might need to layer on more tools. (That said, I once tried integrating it with a computer vision project, and it was smoother than expected, even if I had to tweak a few lines.)

One thing that stands out is how Wool handles fault tolerance—it's designed to keep things running if a node fails, which is crucial in AI where experiments can crash and burn unexpectedly. And while it's not revolutionary, I believe it fills a gap for developers who want something simple without the corporate baggage of bigger systems. In my view, that's what makes it appealing for the AI community today; it's about getting work done faster, not reinventing the wheel.

All right, let's wrap this up with a quick look at where Wool fits in the broader picture. For AI enthusiasts, it's a tool that could streamline workflows, especially if you're into machine learning experimentation. I think it'll gain traction, but only if the community chips in with more examples and fixes. Oh, and speaking of that, have you ever wondered how these runtimes evolve—wait, maybe that's for another time.

Why Wool Stands Out

It's all about efficiency in distributed setups, which I've tested with some AI benchmarks. This means faster iterations on projects, like when I was building a simple generative AI demo last week.

Potential Drawbacks

Not everything's rosy; integration can be tricky if you're not careful, as I found out the hard way.

The Bigger AI Impact

For machine learning pros, Wool could mean less downtime and more innovation, but it's still early days.

Look, if you're into AI, what do you think about tools like this? Share your experiences in the comments—maybe you've got a story about distributed Python that could help others out.

FAQ

What is Wool exactly?

Wool is a lightweight runtime for running Python code across multiple machines, making it easier for AI tasks without heavy setups.

Is Wool good for beginners in machine learning?

Yeah, it's straightforward, but you might need some Python basics first to get the most out of it.

How does Wool compare to other tools?

It's lighter than Ray, which is great for simple projects, but for complex AI work, you might prefer something more feature-rich.