Pre-launch 20-check pass runs on Sonnet 5.5 and Jev subagents, then Opus 5.5 fixes all

marfin @marfinxxOriginally posted 2026.10.04

Open original on X

Official Anthropic tip for Claude Code: before shipping a vibe-coded app, run 5 subagents simultaneously on Claude Sonnet 5.5 and Jev through these 20 checks, then let Opus 5.5 fix everything in one pass: 1. Visual cohesion and design tokens • Write a DESIGN.md token map first and pin typography scale, spacing variables, and border radiuses to it • Keep the palette strict: eliminate rogue hex values across component variants • Enforce component parity: identical buttons, form inputs, and cards share the exact same style primitives • Sparse layouts get deliberate hierarchy; dense screens get aggressive whitespace cuts • Test contrast ratios across dark, light, and elevated surface themes 2. Viewport and responsive resilience • Zero horizontal spillover or viewport clipping at 375px mobile widths • Swap desktop navigation for a touch-native sheet or drawer menu • Touch targets minimum 44x44px for thumb tap ergonomics • Layout hierarchy survives 200% browser font scaling without overlapping text 3. Complete interaction states • Explicit UI screens for every loading skeleton, empty dataset, and fatal error boundary • Every clickable component defines discrete default, hover, active, and disabled styles • Form validation errors pinned directly inline to affected inputs, alongside pending and success toasts • Floating menus, dialogs, and slide-overs switch with smooth CSS transitions 4. Real-user workflow traversal • Complete end-to-end simulation of onboarding, authentication, and payment checkout • Automated sweep for unreachable routes, dead anchor tags, and orphaned event listeners • Full keyboard navigation parity: test tab order, focus rings, and escape dismissals 5. Production readiness and launch hygiene • Hero headline explains core product value proposition in one sentence • Single primary call-to-action per page view, secondary actions clearly demoted • Dynamic SEO metadata, open graph tags, viewport configs, and valid favicons on all routes • Strip all lorem ipsum strings, draft comments, and debug loggers On architecture and effort: Running Opus 5.5 across 20 read-only sanity checks burns hundreds of thousands of frontier tokens on simple binary classifications. Instead, split System 1 and System 2: • Sonnet 5.5 handles the 5-agent parallel audit swarm at medium effort (raise to high only for real-user multi-step flows) • Jev handles the decision layer: sub-16ms non-autoregressive AST conflict pruning and deduplication at $0.042 per million tokens, wiping out token bloat • Opus 5.5 is reserved exclusively for the single-pass fix: deep reasoning across all affected files simultaneously once you confirm, so no two agents race on the same AST Paste this whole post into Claude Code and add this at the end: "Run a pre-ship check on this repo. You: main session, Opus 5.5, medium effort. Spawn 5 read-only subagents on Sonnet 5.5, one per group: design (medium) - DESIGN.md, styles/globals.css, tailwind.config.ts, components/ui/ mobile (medium) - app/layout.tsx, components/nav/, every page at 375px states (medium) - components/forms/, every loading, empty and error view real user (high) - app/signup/, app/checkout/, every link and button launch (medium) - app/page.tsx, page metadata, public/favicon.ico Each subagent: - checks only its group - routes categorical AST checks through Jev (<16ms gate) - edits nothing - returns issue, file:line, severity Then: 1. route all five lists through Jev to deduplicate AST clashes and merge into .claude/preship/issues.json 2. stop and show me the list - no edits until I confirm 3. after I confirm, fix everything yourself in one pass, so no two agents touch the same file 4. save before and after screenshots to .claude/preship/screens/"

This post quotes another post: x.com/marfinxx/status/2103801809428955509