Elephant Goldfish (EGM) vs a plain fresh review: website test files
====================================================================

The briefs, setup and results from Luke Byrne's video "Does Elephant Goldfish Add Anything to a Fresh Review?"
(Luke Byrne | AI & Tech, @lukebyrnetech). Free, no email needed. Guide: https://lukejbyrne.com/library/

WHAT WAS TESTED
1. Sharing feature (main comparison, 2 Nov 2026). Two desktop Claude Code sessions started from identical website
   files with the same task and the same model settings (Opus 5.5, Medium, Auto permission mode, Fast mode off).
   - Plain review: the prompt in briefs/1-sharing-feature-prompt.txt plus exactly the sentence in
     briefs/2-plain-review-sentence.txt.
   - EGM: the same prompt, with the Elephant Goldfish plugin installed and its review loop run once
     (a comprehension reviewer and an implementation-readiness reviewer).
   Both used two fresh reviewers for one pass, amended the plan, then built. Both passed 11/11 independent checks.
   In the delayed clipboard test (start sharing Tripo, change the search to Codex, fail the old copy request) the
   plain-review build showed the stale Tripo link under the Codex results; the EGM build ignored the old response.
   The plain version's first plan had a guard for this; one of its reviewers suggested cutting it and the main
   agent accepted (sharing-test/plain-review-run/review-2-scope-ux-accessibility.md and revised-plan.md).
   EGM's implementation-readiness reviewer flagged the race (sharing-test/egm-run/reviewer-implementation-readiness.md).
2. Homepage (earlier Codex comparison, 1 Nov 2026). EGM against a Codex agent working normally (no review
   instruction), not against the one-sentence review. Brief: briefs/3-homepage-brief.txt. Both passed 10/10 checks;
   screenshots of both at desktop (EN), tablet (ES) and phone (FR) are in homepage-test/.

SETUP
- Install EGM in Claude Code:  claude plugin marketplace add google/elephant-goldfish-model
                               claude plugin install egm@egm        (then start a new session)
  In the desktop app: Customize > Plugins > add the google/elephant-goldfish-model marketplace and install
  Elephant-Goldfish Model. Codex CLI: codex plugin marketplace add google/elephant-goldfish-model
  then codex plugin add egm@egm.
- EGM arm: paste the sharing prompt, then run /egm:egm-review-loop on the plan (one pass), amend, build.
- Plain arm: paste the sharing prompt with the plain sentence at the end. Tell it not to use EGM.
- Keep each run in its own copy of the project, on a local branch, away from the live site and from each other.
- Run the harness against each finished build: node evaluate.cjs / evaluate-semantic.cjs (11 checks) and
  node clipboard-race.cjs (the delayed clipboard case). They were written for the lukejbyrne.com library code
  (jsdom); adapt the selectors and paths to your own project.

LIMITS
One paired run per task, not equal-token, cost-matched or repeated trials. The delayed clipboard failure was
reproduced deterministically in jsdom; how often it happens in a real browser was not measured. The homepage designs
were already decided before EGM's reviewers saw the plan, so a preference for the EGM design is not credited to the
reviewers. Google's repository says EGM is not an officially supported Google product.

NOT INCLUDED
- The full website patches (the builds touch the whole site source; left out to keep this small and free of
  unrelated site code and configuration).
- Local machine paths and session IDs, replaced with <local path>, <checkout> or <session>.
