An AI-resilient workflow stays reliable when a model or provider changes. Treat AI as a useful but inconsistent component, and verify its output against stable requirements instead of depending on predictable behavior. section 14.5
Keep the specification as the stable definition of correct. Put reusable prompts, skills, or project rules between the spec and the AI; when behavior changes, adjust these artifacts and check that results still meet the spec. section 14.5
After a model change
If prompts that once produced focused implementations now produce sprawling ones, avoid piling on one-off constraints. Test saved templates or skills against the unchanged spec, then revise the reusable instructions as needed. section 14.5
Make the workflow adjustable:
- Promote prompts that work repeatedly into saved skills, project rules, or templates. section 14.5
- Version prompt templates in the repository so they can be reviewed and tuned when AI behavior changes. section 14.5
Figure 14.4 The spec as the stable anchor in an AI-assisted workflow: the AI varies, the acceptance criteria don't.

The diagram highlights the key separation: AI behavior varies, while the spec and acceptance criteria remain the anchor. section 14.5
