CHAPTERS
- 0:04 – 1:04
New releases: OpenAI Codex app + GPT-5.3 Codex vs Anthropic Opus 4.6 (and Fast)
Claire frames the episode around two major AI coding releases from OpenAI and Anthropic and sets expectations for a hands-on, side-by-side comparison. She teases a big productivity jump from using these tools while noting each model has quirks and failure modes.
- •OpenAI’s Codex desktop app and GPT-5.3 Codex model release
- •Anthropic’s Opus 4.6 and Opus 4.6 Fast release
- •Side-by-side evaluation promise: where models shine vs fall apart
- •High velocity shipping claim as the backdrop for the review
- 1:04 – 2:05
Sponsor break: Why enterprise AI apps need auth, controls, and auditability (WorkOS)
The sponsor segment explains that AI tools only work well with deep access to company systems, which increases security and compliance demands. WorkOS is positioned as a drop-in solution for enterprise-ready features so teams don’t build them from scratch.
- •AI apps require deep access to codebases and internal docs to be useful
- •Enterprise buyers demand security: auth, access controls, audit logs
- •Building enterprise features in-house is costly and slow
- •WorkOS offers APIs to accelerate enterprise readiness
- 2:05 – 2:35
Test methodology: pick a hard, comparable task inside a real codebase
Claire explains how she evaluates models: choose an ambitious task she wouldn’t do manually, and run the same challenge across models for a fair comparison. Instead of a one-shot landing page, she uses an established site with multiple pages and real complexity.
- •Use one consistent task to compare models side-by-side
- •Avoid simplistic “one-shot landing page” tests
- •Prefer established codebases to reveal real strengths/weaknesses
- •Goal: meaningful comparison of autonomy and quality
- 2:35 – 3:36
The benchmark task: redesign ChatPRD marketing site for PLG + enterprise
She describes the current site and why it needs an enterprise polish while staying friendly for product-led growth. The prompt is intentionally high-level to test whether models can plan, execute, and maintain coherence across a multi-page redesign.
- •Target repo: ChatPRD marketing site with many pages/blog/workflows
- •Desired repositioning: PLG-friendly but enterprise-ready
- •High-level prompt to test autonomy and long-running execution
- •Models tested in release order, starting with Codex
- 3:36 – 9:09
Codex desktop app walkthrough: Git-native workflow, skills, and automations
Before coding, Claire highlights what makes Codex unique: Git primitives are first-class, plus built-in concepts for reusable “skills” and scheduled “automations.” She argues this UI can help both power users running parallel agent work and newer engineers learning Git concepts.
- •Projects map to Git repos; Git concepts are central in the UI
- •Branches vs worktrees for parallel agent work
- •Diff and PR creation built into the workflow
- •Skills as packaged prompts/instructions with a visual UI
- •Automations: scheduled prompts with recommended templates
- 9:09 – 15:14
Running the redesign in GPT-5.2 Codex: strong coding, but overly literal and hard to steer
Claire runs the redesign using GPT-5.2 Codex (noting 5.3 arrived shortly after) and quickly hits a recurring issue: the model follows prompts too literally and overfits to the latest instruction. The back-and-forth makes broad, creative work feel slower despite the model being fast.
- •Test executed on GPT-5.2 Codex due to timing; expects similar family behavior
- •Codex is “too literal,” turning instructions into blunt copy and structure
- •Fails to naturally push toward next steps (planning/build readiness)
- •Overfits to the most recent prompt, losing balance and nuance
- •Example failure: misinterpreting “content-dense” into “dense workflow” headline
- 15:14 – 16:45
Codex results: acceptable homepage/enterprise refresh, but limited scope vs request
She reviews the before/after and finds the output visually “okay” with some good headlines and graphics usage, but not aligned with her desired aesthetic. More importantly, Codex didn’t follow through on the full-site redesign and mostly updated only a couple pages.
- •Output is not terrible and likely faster than doing it manually
- •Some copy/headlines and graphic pulls are decent
- •Design doesn’t fully match brand aesthetic; placeholders remain
- •Major gap: only redesigned homepage and enterprise page
- •Conclusion: can do quality work, but broad site-level execution lagged
- 16:45 – 19:18
Switching to Opus 4.6 in Cursor: better planning and long-running execution (with an early design stumble)
Using the same prompt, Opus 4.6 shows stronger autonomy: it explores the codebase, creates a plan (via Cursor Plan mode), and executes more end-to-end. The first iteration looks unsophisticated, but after targeted feedback it rebuilds with a more distinctive, polished style.
- •Opus 4.6 plans and executes long tasks more independently
- •Cursor’s harness (plan/to-dos/exploration) may boost outcomes vs Codex app
- •Initial result: great copy, poor visual design (‘Tailwind Indigo slop’)
- •After feedback, it rebuilds into a modern, unique visual style
- •Integrates design system and improves cohesion across pages
- 19:18 – 21:19
Opus 4.6 final redesign: cohesive brand, enterprise-friendly value props, and broader site coverage
Claire shows the improved before/after and is enthusiastic about shipping it. Opus keeps the brand feel while making the site more premium, using real assets, stronger value propositions, and extending the redesign consistently across additional pages like pricing and How I AI.
- •Polished visuals while retaining brand colors and personality
- •Uses repo graphics instead of placeholders; adds metrics/social proof
- •Enterprise messaging becomes value-prop oriented, not blunt labels
- •Successfully extends styles across the rest of the site
- •Assessment: strong for generative, greenfield, design-heavy work
- 21:19 – 22:50
Real-world engineering impact: 44 PRs in five days by pairing Opus (builder) + Codex (reviewer)
To move beyond the marketing-site demo, Claire shares concrete shipping stats from her core application work: dozens of PRs, thousands of lines changed, and multiple integrations. She introduces a complementary workflow where Opus drives implementation and Codex hardens it through rigorous review.
- •44 PRs, 98 commits, 1,088 files touched in five days
- •~93k lines added / ~87k removed; ~5k net new
- •Shipped multiple MCP integrations and major component refactors
- •Models used together: Opus for building, Codex for critique and edge cases
- •Framing: assessing real engineering usefulness beyond visual demos
- 22:50 – 24:21
Component refactor case study: Opus 4.6 plans and rebuilds a messy tool-components system
Claire describes refactoring a set of inconsistent MCP/tool UI components into a cleaner, reusable system. Opus 4.6, guided by Cursor Plan mode, produces an extensible component approach with improved consistency across different tool-call experiences.
- •Problem: many tools displayed differently; messy, duplicated components
- •Goal: reusable, customizable shared component architecture
- •Workflow: Cursor + Opus Plan mode, iterative implementation
- •Outcome: extensible tool component with better UI consistency
- •Opus excels at getting to an 80–90% complete implementation quickly
- 24:21 – 26:53
Where Codex wins: architecture/performance review, edge cases, and “principal engineer” feedback
Claire brings the refactor into Codex for pre-ship review, asking for architecture and performance feedback. Codex identifies high-impact issues, prioritizes them, asks clarifying questions, and helps implement polish—reinforcing her view that Codex is exceptional at critique and hardening.
- •Codex used as reviewer: architecture, performance, scalability concerns
- •Finds and prioritizes high-impact issues; probes with questions
- •Implements polish fixes after alignment on intent/edge cases
- •Pairs well with Bugbot/Codex-based review tooling
- •Characterization: Codex as ‘principal engineer’ who critiques more than creates
- 26:53 – 28:55
Cost and tradeoffs: Opus 4.6 Fast speed vs budget realities
Claire briefly evaluates Opus 4.6 Fast, emphasizing it’s significantly more expensive and should be matched to the right tasks. She argues the ROI can still be strong given the throughput gains, but warns against accidentally using premium modes on the wrong work.
- •Opus 4.6 Fast offers higher speed/power but at ~6x cost
- •Approximate pricing discussed per million output tokens
- •Token spending can be high; requires active budget awareness
- •ROI argument: shipping velocity can outweigh model costs
- •Advice: choose model tier based on task value and cost sensitivity
- 28:55 – 30:12
Final verdict: Opus for creative build work, Codex for hardened engineering review—and a multi-model stack
Claire summarizes the head-to-head: Opus 4.6 is her go-to for creating and redesigning, while GPT-5.3 Codex excels at bug-finding, architecture review, and producing robust code. She closes with where these tools fit in her stack and invites viewers to share their experiences.
- •Opus: feature/product work, creative generation, high-quality design execution
- •Codex: bug-catching, architectural analysis, hardened code quality
- •Best results come from pairing: Opus builds, Codex critiques
- •Tooling preference: Cursor remains her primary harness
- •Wrap-up and call to subscribe/comment
