Your approach to code/develop with CC or other "harnesses"
retained text
Hi everyone, I’m constantly exploring the AI tooling space, especially from the perspective of a software engineer working with real-world codebases. One thing I keep noticing is that many of the popular agentic development workflows and methodologies - things like BMAD, GSD, various skill-based setups, prompt frameworks, etc. - feel limited once you move beyond simple greenfield projects. They look great when the task is something like: build a new e-commerce application, create a new SaaS portal, implement a feature in small repo. But I’m not convinced they translate equally well to mature brownfield systems. In larger projects you often have multiple services, existing infrastructure, years of architectural decisions, conventions, technical debt, deployment constraints, security rules, domain-specific assumptions, and different technologies living next to each other. That is where many of these workflows start to feel too generic. For example, I don’t think a generic skill should behave the same way for: * a Java codebase with banking background where consistency is most important, * a Gleam application for small backend, * a React app with ui/ux as the most crucial part, * Terraform/Kubernetes infrastructure , * or a performance-sensitive streaming service. The engineering principles, tools, validation mechanisms, testing strategies, architecture expectations, and even the definition of "good code" can be very different. Another thing that bothers me is determinism. A lot of these workflows are primarily collections of .MD files and prompts interpreted by an LLM. That means a large part of the workflow itself is still probabilistic. Personally, I’d prefer a stronger separation between: LLM-driven reasoning and deterministic workflow enforcement. For example, I’d like the model to reason about implementation or architecture, while a CLI/runtime/hooks enforce things such as: * workflow state, * required steps, * dependencies between tasks, * validation commands, tests, linters * approvals, * evidence of completion, * architectural constraints, * allowed/forbidden operations, So instead of the model being told: "Remember to run tests and review the implementation" the workflow deterministically require those things before the task can be considered complete. What I’m looking for is something closer to an extensible engineering harness for coding agents rather than another prompt framework, but without all these boilerplates, tons of UI, TUI, GUI and floppy features. Ideally it would have a small generic core, but allow projects to define their own behaviour. Something like: `core workflow engine` \- `ecosystem-specific rules/skills` \- `repository-specific rules/skills` \- `task-specific workflow` For example, a Java project could provide Java/JVM/Spring-specific implementation and review rules, while a Rust, Gleam, Go, frontend or infrastructure project could plug in completely different ones. As for example for backend crucial tasks would be great to have some evidences of ADR or small diagrams and for UI to have Figma's design before writing a code. Another important part for me would be some form of project memory/learning. During development, the system should gradually understand and preserve things like: * ADRs, * conventions, * deployment assumptions, * known architectural boundaries, * mementos. Continuously building structured knowledge about the repository and using it during future work. Something that provides enough structure to make agent behaviour predictable, but doesn't try to encode all software engineering into a few generic MD files/prompts or something else. I’m currently experimenting with building something like this myself, but it’s still far from being mature enough for regular use. So my question is: Are there any projects, frameworks, research efforts, or engineering workflows already moving in this direction? To be at once simple, but consistent and flexible (maybe pluginable?). I’d be very interested in seeing what do you use in your work.
document fetched 2026-09-20 00:36 · canonical https://reddit.com/r/ClaudeCode/comments/1wkpeu5/your_approach_to_codedevelop_with_cc_or_other