yaml-config

YAML / Config / Infra Rules

Universal

• Match existing indentation exactly. YAML is whitespace-sensitive — never change indent width.
• Match existing key ordering; don't sort alphabetically as a side-effect.
• Match existing quoting (none / single / double).
• Show only the changed key(s) and the minimum surrounding context to anchor location.

Output

• For a single key change, show the parent path + the changed line:

  spec:
    template:
      spec:
        containers:
          - name: api
            image: myrepo/api:v1.4.2  # ← changed
  

• Don't reprint the entire manifest.

Dockerfile

• Match existing base image strategy; don't switch from alpine to debian unless asked.
• Don't reorder layers — that invalidates caches the user is depending on.
• Don't combine/split RUN instructions unless that's the task.
• Don't add HEALTHCHECK / USER / LABEL unless siblings have them.

docker-compose

• Match existing version key.
• Don't add new services to a feature task — only what's asked.
• Don't change exposed ports unless requested.

Kubernetes / Helm

• Don't change apiVersion unless that's the explicit task.
• Don't add new resources (HPA, PDB, NetworkPolicy) speculatively.
• Match existing label/annotation conventions exactly.
• For Helm: don't move values between values.yaml and templates without flagging.

Terraform

• Match existing module structure — don't refactor into new modules.
• Don't change provider versions while doing a feature.
• Don't add new providers.
• Match existing variable naming (snake_case is conventional).
• Show only the changed resource / module block.

GitHub Actions / CI

• Match existing workflow trigger style.
• Don't add new actions/marketplace dependencies without flagging.
• Don't pin/unpin action versions as a side-effect.
• Don't add caching steps to workflows that don't have them — that changes runtime behavior.

Never

• Don't add comments to production manifests unless a value is non-obvious.
• Don't introduce environment variables that need new secrets without flagging.
• Don't change resource limits/requests as a side-effect.
• Don't suggest "switch to a different infra tool" as a fix.