Demonstration · 15 Nov 2026

React Best Practices

React and Next.js performance guidelines from Vercel Engineering: 70 rules in 8 categories, ranked by impact.

A demonstration Techtree ran itself on Vercel’s public React Best Practices Skill, from its agent-skills collection (MIT license), used unchanged as the starting point. It is not customer work.

Task

Given a short piece of React, Next.js or TypeScript code and five of the Skill’s rule ids, reply with the id of the one rule the code breaks, and nothing else; an answer counts only when it matches exactly.

Result

The final test: 37 cases nothing had seen, each asked 3 times of each version, 111 attempts per side.

Original 54% 60 of 111
Revised 51% 57 of 111
Verdict
Inconclusive
Difference
−3 attempts
Cases
9 better, 19 the same, 9 worse
Chance
About 74.5 in 100

The change did not help. The plain rewrite solved 3 fewer attempts than the original. The test did not establish a difference between the two versions. If the two versions were equally good, a difference this large would turn up by chance about 74.5 times in 100. The line agreed before anything ran is 5 in 100, so the test cannot rule out chance.

Method

  1. Split

    We wrote 83 cases for this test and split them once, before anything ran: 25 to work on the Skill, 21 to choose between versions and 37 kept apart for the final test. Every version was asked each case 3 times.

  2. Improve

    Two revisions were made from the 25 working cases alone: a single plain rewrite by Anthropic’s Claude Opus 5.5, and a revision by Microsoft’s SkillOpt, with Anthropic’s Claude Opus 5.5 proposing its changes. On those cases the agent solved 53% of its attempts with no Skill (40 of 75) and 69% with the original (52 of 75).

  3. Choose

    The original and both revisions were tried on the 21 choosing cases. The plain rewrite solved the most and became the improved Skill.

    Original 71% 45 of 63
    Rewrite 73% 46 of 63
    SkillOpt 59% 37 of 63
  4. Final test

    One run on the 37 cases nothing had seen, 111 attempts per version, with the same agent and settings on both sides.

Cases

9 cases did worse with the improved Skill: billing-email-action went from 2 of 3 to 1 of 3; comment-emoji-data went from 3 of 3 to 2 of 3; data-table-hidden-columns went from 2 of 3 to 1 of 3; discount-rate-loop went from 2 of 3 to 1 of 3; drag-handle-pointer went from 2 of 3 to 1 of 3; pages-app-recorder went from 2 of 3 to 0 of 3; photo-editor-click went from 2 of 3 to 0 of 3; team-lookup-object-arg went from 2 of 3 to 0 of 3; viewer-lookup-per-request went from 3 of 3 to 0 of 3.

Case Original Revised Change
activity-feed-items 3 of 3 3 of 3 0
archive-project-action 0 of 3 0 of 3 0
avatar-menu-account 3 of 3 3 of 3 0
back-to-top-scroll 1 of 3 1 of 3 0
billing-email-action 2 of 3 1 of 3 −1
catalog-grid-products 3 of 3 3 of 3 0
comment-emoji-data 3 of 3 2 of 3 −1
data-table-hidden-columns 2 of 3 1 of 3 −1
digest-dry-run 0 of 3 2 of 3 +2
discount-rate-loop 2 of 3 1 of 3 −1
draft-save-handler 3 of 3 3 of 3 0
drag-handle-pointer 2 of 3 1 of 3 −1
editor-toolbar-icons 1 of 3 3 of 3 +2
inbox-empty-illustration 0 of 3 0 of 3 0
inbox-messages-wait 1 of 3 2 of 3 +1
layout-speed-insights 1 of 3 2 of 3 +1
log-filter-typing 0 of 3 1 of 3 +1
map-wheel-hint 2 of 3 3 of 3 +1
orders-page-layout 1 of 3 2 of 3 +1
pages-app-recorder 2 of 3 0 of 3 −2
photo-editor-click 2 of 3 0 of 3 −2
place-order-effect 2 of 3 2 of 3 0
project-activity-wait 1 of 3 1 of 3 0
purge-cdn-regions 3 of 3 3 of 3 0
ranked-players-prop 3 of 3 3 of 3 0
refund-order-action 0 of 3 0 of 3 0
rename-form-saving 1 of 3 1 of 3 0
submit-disabled-memo 1 of 3 2 of 3 +1
sync-icon-spin 3 of 3 3 of 3 0
sync-settings-controls 0 of 3 0 of 3 0
tag-selection-compare 1 of 3 1 of 3 0
team-lookup-object-arg 2 of 3 0 of 3 −2
tenant-module-variable 2 of 3 3 of 3 +1
ticket-counts-filters 3 of 3 3 of 3 0
viewer-lookup-per-request 3 of 3 0 of 3 −3

Tools

Agent
NVIDIA’s Nemotron 3 Nano 4B (nemotron-3-nano:4b), run on our own machine.
Scoring
Prime Intellect’s Verifiers 0.4.1, with the same settings on both sides.
Check
NVIDIA’s SkillEvaluator 0.9.0 reviewed the original Skill as written.
Rewrite
Anthropic’s Claude Opus 5.5, one plain rewrite from the working cases.
Revision
Microsoft’s SkillOpt, with Anthropic’s Claude Opus 5.5 proposing its changes, on the working cases only.
Skill
Vercel’s React Best Practices Skill, the starting point, MIT license.

Report

The report lists every final-test case with the attempts each version solved, so the figures above can be counted again from it.