[brentvatne]

Why

#86 added a rule for rating design findings to templates/shared.md and templates/agents/correctness.md, the files adopters get. This repo reviews its own PRs with its own copies in .expo-agents/code-review/, and those didn't get it.

Without the rule, High and Medium confidence are defined only by "the failure or exploit", and Low findings are not reported. A design finding, which has a real cost but no failure path, rates Low and the coordinator drops it. The AI review of #87 flagged this (id:ae47861897b3).

What

The same two lines as the templates:

• shared.md: a design finding's confidence is how directly the code shows its cost, and its impact is how large that cost is.
• agents/correctness.md: "no concrete failure path (or, for a design finding, no concrete cost)".

This only changes how this repo reviews itself. The published package doesn't change.

Test plan

ecr ref-check and ecr verify-config pass on this repo's setup.

🤖 Generated with Claude Code

[github-actions]

🤖 AI code review

Decision: Ready for human review

Overall PR risk: Low. This additive change updates internal reviewer prompts to match published templates. No runtime code or workflows change.

🔎 Reviewer summaries (3)

• consistency: I checked both prompt edits against the scaffolding copies in templates/. templates/shared.md and templates/agents/correctness.md already have the same design-finding text. I found no consistency problems.
• correctness: I reviewed the prompt-only edits that define Confidence and Impact for design findings. The edits match the wording in templates/ and agree with the severity definitions. I found no issues.
• security: Reviewed two reviewer-prompt Markdown edits that define how design findings rate Confidence and Impact. The edits add no code, secrets, shell use, or change to the injection boundary. No security findings.

No findings.

---

This review is advisory — it never blocks a merge and never auto-approves.