How we build products
CHEEE combines product direction, design and frontend engineering with AI-assisted development. We start from a clear user need, define what the product should do, then build, check and refine the working application. AI tools accelerate the work; JR and SP own the decisions and quality.
Start with the task.
Build a useful first version.
For workflow tools, that means looking at the information people handle, the steps they repeat, and the moments where work moves from one person to another. For a campaign, it means getting clear on the audience and what they should learn, try or take part in.
Claude, Codex and OpenCode support code exploration, prototyping, implementation and debugging. We review their output in the context of the product, test how it works and keep important decisions visible.
-
Understand the task
Learn how the work happens today, who is involved and where it gets difficult. We look at the information people start with, the steps they repeat and what happens when a task needs an exception. For a campaign, start with the audience and what they should learn or take part in.
-
Define the product clearly
Focus on one repeated task or one player goal. Decide what the product must help someone do, which details matter to the experience and what can wait. Clear priorities keep the product useful and cohesive.
-
Build around the real flow
Fit the tool or experience to the steps people actually take. If AI is proposed as an in-app feature, give it a defined job and make a person's review or decision points clear. People should understand what the feature did and what still needs attention.
-
Refine what people need
Use what is learned from the first version to improve the steps and interaction. Look for what is confusing, what is missing and what needs to happen next. Keep the parts that help, change the parts that slow people down, and let evidence from use guide the next decision.
Tools chosen
for the task.
We choose from a broad set of AI and development tools to fit the product and the task. The value is in the application or game we deliver, not the number of tools involved.
Claude is part of our development workflow, alongside Codex, OpenCode and Orca. We use AI where it helps us work through product questions, create software and make media assets.
-
AI-assisted product development
Claude, Codex and OpenCode help with code exploration, interface prototyping, implementation, debugging and iteration. We select the right assistant for the task and review its output before it becomes part of the product.
-
Cross-agent orchestration
Orca coordinates work across agents when parallel work can move the product forward. People still integrate, review and take responsibility for the result.
-
Media assets
Nano-Banana-2.1, GPT-image, Seedance and MiniMax-H3 support image, motion and other media work. We choose based on the asset, format and control the product needs.
AI can assist.
People stay accountable.
“AI-assisted” describes how we make software. “AI-powered” describes something the finished application does while a person is using it. We keep those claims separate.
We define the problem and product direction, make experience and architecture decisions, review generated work, check the functioning interface and decide what is ready to share. The tools can help us get to a draft sooner; they do not get the last word.
-
Problem and product direction
Choose what the project is for, who it needs to help and what a useful outcome looks like.
-
Design and implementation
Shape the interface and technical approach, then check how the working version behaves across real screens and interactions.
-
Review and release
Inspect AI-assisted output, test the result and make the final call about what should be shared.
Where AI can add product value
We see opportunities for model-powered features where they improve the job. The capabilities below are future product directions, not current demo features; each will need a clear user benefit, human review and sensible limits.
-
Make website references easier to use
A Claude-powered feature could turn website structure and content into concise, reviewable design or implementation notes. Site Stripper already gathers reference material; adding model-generated analysis is a future product direction, not a feature in the current demo.
-
Turn loose notes into a useful first draft
AI-assisted applications can organise project notes or requests into structured drafts that a person can inspect, correct and use as a starting point.
-
Add context to an interactive experience
Selected model-assisted responses could bring more context to a story or game when they improve the interaction and remain understandable. None of the games shown here is presented as using a language model.
Describe the task or campaign.
A useful first note can be simple: the task, where it slows down and what a better finish would look like. For campaign work, tell us who it is for and what people should be able to learn or do.