asnowwolf/meta-project-poc

Meta Project Demo is a proof of concept for organizing how a person uses large language models to build software, rather than being a tool that writes code by itself. The idea behind it is that a single open ended conversation with an AI, relying only on whatever fits in its context window, tends to lose track of decisions, produce inconsistent quality, and drift away from the original goal as a project grows. This project tries to fix that by borrowing established software engineering methods and turning the overall process, not the code itself, into something structured and repeatable. The workflow moves through a sequence of stages that mirror real development practice: design thinking for figuring out user needs, domain driven design for modeling the business concepts involved, an architecture stage for deciding how the system is structured, agile planning to break work into sprints, and test driven development for the actual coding, followed by a retrospective stage for reviewing what happened. Each stage produces a concrete document or artifact, such as a requirements write up or a domain model, which becomes an input to the next stage. A key design choice is that the big picture decisions, which methodology to use, what order stages happen in, which AI model handles which stage, are fixed in advance by the person running the process or by a script, while the fine grained details, like exactly how a class is implemented, are left for the AI to work out within those boundaries. Certain points in the process are marked as human checkpoints, where the person running it must review the output before the process continues, and these checkpoints can be configured to always pause, never pause, or only pause under certain conditions. Each stage also repeats automatically, generating output, checking it against a quality bar, and trying again with feedback until it passes or reaches a limit, at which point it waits for a person.