Use Opus and GPT in reviewer docs

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: richlander

.github/skills/release-notes/DESIGN.md

@@ -44,13 +44,13 @@ The system still has three layers, but the AI/editorial layer is now split into
 - **`generate-features`** adds reusable triage and scoring so release notes, docs, and blogs can share the same ranked feature list.
 - **`api-diff`** verifies the public API story against the actual build, which is the best defense against missed reverts.
 - **`release-notes`** focuses on editorial judgment: which scored items matter, how to describe them, and how to preserve human edits.
-- **`review-release-notes`** applies a harsher, reader-centric cutoff so the final draft does not drift into API-inventory territory. The final pass should fan out to three different models so the system gets a broader editorial read before synthesis.
+- **`review-release-notes`** applies a harsher, reader-centric cutoff so the final draft does not drift into API-inventory territory. The final pass should fan out to two different models so the system gets a broader editorial read before synthesis.
 - **The workflow** handles orchestration. It knows when to run, what branches to manage, how to interact with humans, and how to preserve their edits.
 
 ### Model strategy
 
-- **Goldeneye by default** — use Goldeneye throughout the main orchestration, scoring, and drafting flow.
-- **Three-model final review** — the preferred final editorial trio is **Goldeneye + Claude Opus 4.6 + GPT-5.4**.
+- **Claude Opus 4.6 by default** — use Claude Opus 4.6 throughout the main orchestration, scoring, and drafting flow.
+- **Two-model final review** — the preferred final editorial pair is **Claude Opus 4.6 + GPT-5.4**.
 - **Synthesis over voting** — overlap is a strong signal, but the shared rubric and fidelity to the actual shipped changes still outrank raw vote count.
 
 ## Layer 1 — The Tool

.github/skills/release-notes/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: release-notes
 description: Generate and maintain .NET release notes from `features.json`. Uses `generate-changes` for authoritative shipped-change data, `generate-features` for scoring/triage, `editorial-scoring` for the shared rubric, `api-diff`/`dotnet-inspect` for API verification, and a multi-model `review-release-notes` pass for final editorial QA.
-compatibility: Requires GitHub MCP server or gh CLI for cross-repo queries. Pairs with the generate-changes, generate-features, editorial-scoring, api-diff, and review-release-notes skills. Goldeneye is the default workflow model; the preferred final reviewer trio is Goldeneye + Claude Opus 4.6 + GPT-5.4 for broader editorial feedback.
+compatibility: Requires GitHub MCP server or gh CLI for cross-repo queries. Pairs with the generate-changes, generate-features, editorial-scoring, api-diff, and review-release-notes skills. Claude Opus 4.6 is the default workflow model; the preferred final reviewer pair is Claude Opus 4.6 + GPT-5.4 for broader editorial feedback.
 ---
 
 # .NET Release Notes

.github/skills/review-release-notes/SKILL.md

@@ -78,15 +78,14 @@ Compare the draft against the component examples:
 
 ## Multi-model review pattern
 
-For the final editorial QA pass, use this skill as a **three-reviewer parallel check** to get broader viewpoint diversity:
+For the final editorial QA pass, use this skill as a **two-reviewer parallel check** to get broader viewpoint diversity:
 
 Preferred set:
 
-1. **Goldeneye**
-2. **Claude Opus 4.6**
-3. **GPT-5.4**
+1. **Claude Opus 4.6**
+2. **GPT-5.4**
 
-Give all three reviewers the same inputs and the same requested output:
+Give both reviewers the same inputs and the same requested output:
 
 - most correct scoring
 - most wrong scoring

.github/workflows/release-notes.md

@@ -132,12 +132,11 @@ Read these files and skills for detailed guidance:
 
 ## Model strategy
 
-- Use **Goldeneye** as the default model throughout the workflow for orchestration, scoring, and drafting.
-- For the **final `review-release-notes` pass**, prefer this **three-model reviewer set** to widen the editorial viewpoint:
-  - **Goldeneye**
+- Use **Claude Opus 4.6** as the default model throughout the workflow for orchestration, scoring, and drafting.
+- For the **final `review-release-notes` pass**, prefer this **two-model reviewer set** to widen the editorial viewpoint:
   - **Claude Opus 4.6**
   - **GPT-5.4**
-- Give all three reviewers the same inputs and ask for the same output shape. Synthesize the overlap, inspect meaningful disagreements, and prefer the shared `editorial-scoring` rubric over any single model's preference.
+- Give both reviewers the same inputs and ask for the same output shape. Synthesize the overlap, inspect meaningful disagreements, and prefer the shared `editorial-scoring` rubric over any single model's preference.
 
 ## What to do each run
 
@@ -278,11 +277,10 @@ When unsure about a human's intent, ask. Use `add-comment` to reply. This is a c
 
 #### h. Run the final multi-model review
 
-Before pushing the draft, run the `review-release-notes` stage as a **three-agent parallel review**:
+Before pushing the draft, run the `review-release-notes` stage as a **two-agent parallel review**:
 
-- **Reviewer 1:** Goldeneye
-- **Reviewer 2:** Claude Opus 4.6
-- **Reviewer 3:** GPT-5.4
+- **Reviewer 1:** Claude Opus 4.6
+- **Reviewer 2:** GPT-5.4
 
 Have each reviewer critique the same draft using the same rubric, examples, and
 this checklist: