BobMan
Category: Apps for Your Life
Elevator pitch: BobMan turns family game-making into shared learning: two children are product owners while a software people manager and Codex with GPT-5.6 use a robust harness to build a polished native C++ game.
Inspiration
Most families consume technology together. We wanted to create it together.
BobMan began as a small private C++ game and became a family activity. My children, aged nine and twelve, act as product owners: they propose and prioritize features, define requirements and acceptance criteria, test the results, and decide whether a feature is accepted.
They do not write the code, and BobMan is submitted as my solo project. But they experience a real product-development process through something they care about. They learn that ideas need clear requirements, priorities require trade-offs, and “done” means satisfying observable acceptance criteria.
There is a second personal motivation. I work in software development as a people manager, not as a developer, and have been away from hands-on software engineering for about fifteen years. Small private projects helped me keep in touch, but limited time made it difficult to catch up with modern engineering.
ChatGPT and Codex fundamentally changed that learning curve. They became more than code generators: they provided continuous, practical coaching inside a real and increasingly complex system.
What we built
BobMan is a native C++17 and SDL2 game with animated creatures and gameplay systems, a map editor, persistent configuration, an integrated game guide, automated tests, headless simulation, and macOS packaging.
It is not a thin AI wrapper. It is an asset-heavy application with rendering, audio, persistence, multithreaded behavior, and organically grown complexity.
During Build Week, Codex with GPT-5.6 Sol added dragon gameplay, animated treasure-chest rewards, hierarchical runtime configuration, improved creature and status effects, an icon-based editor workflow, and expanded test and release automation.
How we built it
Our workflow has clear responsibilities:
- My children provide product ideas, priorities, acceptance criteria, and final acceptance.
- I translate their ideas into product and engineering decisions.
- Codex analyzes and implements the approved work.
- Tests, architecture checks, and guardrails verify the result.
- The family reviews the working product.
Over time, I encoded what we learned into the repository. Bugs became regression tests, architectural decisions became enforced module boundaries, and recurring mistakes became automated guardrails. The resulting harness gives agents durable project knowledge while preserving human control over scope, trade-offs, risk, and acceptance.
The GPT-5.6 breakthrough
GPT-5.6 Sol made a qualitative difference in how much responsibility this harness could safely support.
BobMan had accumulated several very large source files. I asked GPT-5.6 to design a modernization concept that would reduce these monoliths without creating unclear dependencies or unnecessary abstraction. It produced a plan and waited for my approval.
After approval, GPT-5.6 executed the refactoring across application orchestration, gameplay, rendering, audio, maps, entities, tests, and documentation without further product guidance during the main implementation. Existing behavior remained intact, and verification did not lead to a user-reported regression or corrective bug-fix loop. Subsequent work came from a deliberate agent-led consistency audit.
This demonstrated more than code generation: a mature project harness enabled an agent to perform complex Brownfield engineering with meaningful autonomy while keeping human decisions and constraints explicit.
Challenges
The hardest challenge was controlling complexity without slowing family experimentation. A visually small request can affect gameplay, rendering, audio, settings, generated animations, and packaging.
Rapid AI-assisted development also creates regression risk. The solution was not increasingly detailed prompting, but teaching the repository to detect mistakes through tests, guardrails, diagnostics, and documented decisions.
Another challenge was learning to use AI without outsourcing judgment. Codex can propose architecture and implement large changes, but I remain responsible for deciding what belongs in the product and whether the result is correct.
What we learned
My children learned that software development is collaborative, not just writing code. They experienced product responsibility, negotiation, testing, and acceptance.
I learned that agentic coding can provide continuous, practical coaching. Instead of studying modern engineering in isolation, I apply it immediately and receive rapid feedback.
This has also made me more effective as a software people manager. I am not trying to become a professional developer again. The technical fluency gained through BobMan helps me make better decisions, ask better questions, understand architectural trade-offs, and help my organization prepare responsibly for agentic development. Colleagues and leaders value this increased depth.
The central lesson is that AI does not have to replace learning. With clear human ownership and a strong engineering harness, it can accelerate learning, deepen judgment, and enable a family to build something meaningful together.
What’s next
BobMan will remain our family product laboratory. My children will continue to shape and prioritize the backlog, while I continue improving the game and its engineering harness.
BobMan is not a game built by AI. It is a family learning how to build together—and a software leader learning how to lead in an agentic future.
Log in or sign up for Devpost to join the conversation.