CHAPTERS
- 0:00 – 2:44
Why Codex beats “just Claude”: the new PM leverage stack
Aakash tees up why Meng To’s perspective matters for PMs: moving beyond basic ChatGPT-style prompting into project-based agent workflows. Meng positions Codex as the center of an ecosystem that can connect to your apps, files, and real work outputs—not just text responses.
- •Codex framed as “ChatGPT on steroids” for real deliverables (slides, HTML, apps)
- •PM value: not just using tools, but using them well (taste, context, workflow)
- •Shift from single-chat usage to connected workflows across apps and local files
- •Claude/Cursor mentioned, but Codex emphasized as the daily driver
- 2:44 – 4:05
Best AI design tools right now (Codex, Cursor, Claude Code, Open-source momentum)
Meng compares the current landscape of AI building/design tools and explains why Codex changed his workflow mindset. He notes rapid convergence as tools copy each other’s best ideas, with Anthropic and OpenAI at the center of the current wave.
- •Codex vs Cursor vs Claude Code: different paradigms, increasingly “project + chat” oriented
- •Open-source tools (e.g., OpenClau) driving local-first workflows and influence
- •Ecosystem arms race is accelerating feature adoption across products
- •Meng’s bias: OpenAI/Codex + Image 2.0 ecosystem for PM-relevant outputs
- 4:05 – 8:43
Codex setup basics: Atlas browser, voice input, and the local-first knowledge base
Meng starts showing the core stack around Codex: an AI browser for contextual querying and agentic browsing, voice dictation for speed, and local note/document systems for organization. The theme is ‘context is king’—capturing and reusing it across projects.
- •Atlas AI browser: ask questions on any page; agent can take over the computer
- •WhisperFlow for fast, accurate voice prompting + custom dictionary
- •Obsidian as a local knowledge base for the flood of AI-generated documents
- •Local-first context improves privacy, capability, and retrievability
- 8:43 – 11:09
How Meng uses Codex projects in real life (folders, outputs, and custom tools)
Meng opens his real Codex workspace and explains how projects map to local folders containing markdown, images, code, and other assets. He also explains why he builds many of his own internal tools: unique workflows require custom solutions, and Codex makes that feasible.
- •Codex ‘projects’ are anchored to real local folders—outputs persist beyond the chat
- •Codex can view/edit MD, HTML, images, diffs; but organization still matters
- •Meng’s philosophy: build tools to solve your own frustrations; AI makes it practical
- •Codex used for prep docs, storyboards, charts, scripts, and production assets
- 11:09 – 13:35
Obsidian’s role + why avatars matter for Loom/UGC communication
Aakash asks for clarity on Obsidian and avatars; Meng explains both as workflow multipliers. Obsidian organizes and links local documents; avatars reduce friction for recording updates and enable a more human, scalable communication style for teams and marketing.
- •Obsidian: organize local docs; graph/‘brain tree’ for long-term retrieval
- •Codex generates many files—Obsidian becomes the system of record
- •UGC shift: audiences prefer authentic creator-style communication over corporate messaging
- •Avatars let you deliver Loom-style updates without perfect lighting/appearance
- 13:35 – 20:04
Plugins vs skills vs computer use (and the trust/security tradeoffs)
Meng distinguishes deeply integrated plugins from lightweight, user-defined skills, then highlights ‘computer use’ as the universal automation layer. He also addresses security concerns from early open-source agents and why guardrails + permissioning matter for trust.
- •Plugins: deeper integrations (Slack, Chrome, Linear, Gmail) vs skills: reusable prompt/tool patterns
- •Computer use works anywhere (acts like a human with mouse/keyboard) vs MCP/CLI requiring integrations
- •Use cases: QA flows, bug reproduction, logged-in navigation, email replies/outreach
- •Security: prompt injection and permissions; Codex seen as more accountable/guardrailed
- 20:04 – 22:00
Building and maintaining your skill library (what to add, when to update)
Meng describes how he updates skills based on what his agents repeatedly misunderstand. He recommends categories of skills depending on what you build—frontend, CSS/Tailwind, deployment, testing, copywriting, and ‘taste’—and encourages PMs/founders to invest in them.
- •Update skills when you notice persistent gaps (e.g., performance, platform specifics)
- •Recommended skill categories: SwiftUI, Playwright, deployment (Vercel/Netlify), copywriting
- •Frontend polish via animation libraries and design-oriented skills
- •PM takeaway: research + integrate skills as your ‘AI operating system’ evolves
- 22:00 – 24:05
Ad break: Arise (trace → evaluate → fix for agent reliability)
Aakash shares a practical loop for improving AI agents: instrument, trace tool calls and decisions, write evals, and iterate. The segment positions evaluation as the missing discipline for builders shipping agentic products.
- •Problem: agents ‘work in testing’ but fail with users due to unseen step-by-step errors
- •Arise installs quickly and auto-instruments LLM/tool-call spans
- •Tracing reveals hallucinations and wrong tool choices; evals quantify failure rates
- •Workflow: trace → evaluate → fix, with fast iteration loops
- 24:05 – 29:37
Codex projects + folder system: the local workspace that makes agents powerful
Meng explains the shift from VS Code/Figma-centric work into project-based chat systems that operate over local folders. He gives a practical folder taxonomy for PMs/founders so agents have enough context without ballooning token costs.
- •Industry shift: from file-centric editors to chat-centric project workspaces
- •A ‘project’ = local folder; create a Projects directory and subfolders per domain
- •Suggested buckets: apps, content/scripts, support/issues, company/admin, skills, even family/school
- •Token management: scope context by folder to avoid excessive/irrelevant context
- 29:37 – 32:41
Permission tiers + Plan Mode: how to start builds safely and efficiently
Meng walks through permission levels (default to full access) and how to choose compute/token settings based on budget and task difficulty. He emphasizes starting with Plan Mode—approve the architecture and steps before letting the agent execute.
- •Trust ladder: default permissions for new users; full access once guardrails are understood
- •Token/compute settings (low→extra high) trade cost/time for accuracy
- •Plan Mode first: get MVP features, architecture, and steps before building
- •Example app idea: QR scanner → email content; ask clarifying questions before coding
- 32:41 – 41:46
Plan Mode in action: building slides from real local data + choosing output formats
Meng demonstrates how to request a slide deck using real metrics stored in project folders, and how to steer deliverables (Keynote vs PPTX vs HTML). He shows how planning, questioning, and correct model/tool naming reduce errors and accelerate execution.
- •Use voice (WhisperFlow) to provide high-context instructions for data pulls and decks
- •Ask explicitly for planning; verify whether a format is supported (Keynote vs PPTX)
- •Steer mid-flight with queued commands; correct tool/model names to avoid misfires
- •Close the loop via HTML (fast/controlled), Figma (slower), or PPTX/Keynote (familiar editing)
- 41:46 – 44:50
The screenshot shortcut and mobile Codex: faster context, always-on execution
Meng highlights a new screenshot shortcut (focused-window capture) as a major workflow unlock—images provide dense context to agents. He also explains Codex Mobile: manage ongoing tasks from your phone while the computer does the work in the background.
- •Screenshots outperform long descriptions: ‘an image is worth a thousand words’
- •New shortcut captures the focused app/window quickly for iterative design feedback
- •Codex Mobile inside ChatGPT app connects to your computer’s running projects
- •Requires computer to stay on + settings/permissions for locked-screen operation
- 44:50 – 54:59
Taste skill: making AI design look senior-level (and where to find it)
Meng explains ‘taste’ as a reusable skill that improves typography, spacing, and visual polish. He shows how to locate taste skills (e.g., GitHub lists) and why HTML-based recreation is often the quickest path to a high-quality, editable output.
- •Taste skills improve baseline design quality but still require human judgment to avoid overcrowding
- •Skills can be sourced from GitHub and dropped in as simple files
- •Experiment with models/settings for speed vs quality (comparisons shown)
- •Workflow: generate images → iterate → recreate as HTML for controllable production output
- 54:59 – 1:05:31
Building your AI digital twin: avatars + video generation tools (and ethics)
Meng outlines a practical pipeline for creating a digital twin using your best photos, then combining it with screen recordings and scripts to produce multi-format content. He recommends specialized tools for avatar/lip-sync video and warns about deception backlash and regulatory pressure.
- •Prompt: select best headshots → generate diverse angles/lighting scenes (Image 2.0)
- •Pipeline: screen recordings + script → avatar overlay → short-form and long-form variants (16:9, 9:16, 4:3)
- •Tooling: HeyGen for presentation-style avatar video; Veo/Omni or others for broader video generation
- •Ethics: disclose/avoid deceptive fake product demos; expect increasing pushback/regulation
- 1:05:31 – 1:12:24
PM survival in layoffs: become ‘technical’ by orchestrating agents, not writing code
The conversation shifts to career strategy: Meng argues non-technical PMs are most at risk, while ‘technical’ now means fluency in AI tooling, jargon, and workflows. He frames the endgame as founding—using AI to handle admin/marketing while you provide domain expertise and quality control.
- •Clarification: layoffs disproportionately hit non-technical PMs; technical PMs persist
- •Technical ≠ coding; it’s understanding models, tools, workflows, and how to direct agents
- •Founder path: AI handles paperwork/accounting/ops/parts of marketing; human provides moat + QA
- •Raising baseline: aim for “11-star” experiences—taste, orchestration, and continuous improvement
- 1:12:24 – 1:14:44
Wrap-up: where to follow Meng + persistence through thousands of prompts
Aakash closes by directing viewers to Meng’s socials and products, emphasizing learning by building. Meng reinforces that high-quality outcomes come from sustained iteration—often thousands of prompts per product—rather than expecting magic from the first attempt.
- •Find Meng on X/LinkedIn/Instagram/TikTok; explore products (Aura, Newform, DreamCut)
- •Reality check: quality takes iteration; don’t quit after the first prompt
- •Aakash’s closing CTAs: subscribe/follow/review/comment + newsletter bundle mention
- •Episode theme restated: PM leverage comes from agent orchestration + workflow systems
