I am specifically interested in hearing from anyone who is working on actual production services, legacy services, etc.

I have done quite a bit of LLM-assisted coding but starting to work towards (and educate others in my org) about more structured approaches to this, to be able to deliver more complicated features with stronger guardrails than prompt-based coding.

Spec-Driven/Anchored development is an emerging practice but I'm not clear on how battle-tested it is. What is your experience with it?

I have used Simon Martinelli's AI Unified Process quite extensively and successfully: https://unifiedprocess.ai

The accompanying tooling allows you to generate requirements for existing projects (brownfield, legacy) as a foundation to either build upon or reimplement existing behaviour.

The whole spec-driven development movement has always been a bit odd to me, because it feels more like a traditional grounded approach towards regular software engineering.

I can share my personal approach which has worked fairly well at least on modest legacy repositories (250k lines of code). It's a pretty simple 8-step process:

Design

• Think about broad project objectives (ABOUT.md)

• Deep research around tooling (STACK.md)

• Initial spec and iterate with LLM (SPEC.md)

• Develop broad suite of tests (TDD.md)

• Full expansion into multi-phase plan with adversarial reviews (PLAN.md)

Execution

• Agentic harness (OpenCode) implementation at a per-phase level

• Green light from TDD and further adversarial reviews. Manual sanity checking also occurs at this stage.

• Repeat implementation, test, validation cycle until feature complete

https://mordenstar.com/other/spec-dev

Something I've learned is that your specification is as solid as your validation harness. If you have added strong e2e tests, with the critical scenarios covered, then you are safe. If not, sometimes it's rolling dices.