CHAPTERS
- 0:00 – 2:44
Why Codex over Claude: the new PM leverage stack
Meng and Aakash frame the episode around a shift from chat-only AI to project-based agent workflows. Meng argues Codex is "ChatGPT on steroids" because it can operate across files, apps, and workflows end-to-end. They set the expectation: this will be a tactical walkthrough aimed at PM use cases, not just prompting theory.
- •Codex positioned as a project-centric AI environment vs standalone chat
- •PMs need workflows that connect to existing apps and local context
- •Meng’s perspective: less Claude, more OpenAI/Codex for daily work
- •Episode goal: go beyond basic prompts into doing the work well
- 2:44 – 8:29
Best AI design + productivity tools (and why context is king)
Meng lists the tools he considers most valuable right now, emphasizing how AI becomes dramatically more useful when it can see your real context. He highlights Atlas (AI browser), WhisperFlow for voice input, and Obsidian for organizing the flood of AI-generated artifacts. He also explains why he builds custom tools to fit unique workflows.
- •Atlas as an AI browser: ask questions and enable agent computer control
- •WhisperFlow for fast, accurate voice-to-text and custom dictionaries
- •Obsidian as a local knowledge base to manage AI-generated documents
- •Why building custom tools matters: every workflow has unique needs
- 8:29 – 11:19
Inside Meng’s real Codex setup: projects, folders, and Obsidian’s role
Meng opens his real Codex workspace and explains how projects map to local folders on his machine. He clarifies that Codex can view many file types, but you still need a durable system for organizing and revisiting outputs—this is where Obsidian and a folder structure come in. The theme is: keep everything local for privacy, control, and better context.
- •Codex projects are backed by local folders; outputs persist on disk
- •Codex can open MD/HTML/images, but retrieval/organization needs a system
- •Obsidian organizes local markdown and builds a linked "brain" of knowledge
- •Local-first context improves power and privacy (influenced by open-source workflows)
- 11:19 – 13:20
Ads break: tracing and evaluating agents with Arize
Aakash explains a common agent failure mode—tool misuse and hallucinations—caused by lack of evaluation. He demonstrates adding Arize tracing via a single command, then using traces to see tool calls and decisions. The key takeaway is the loop: trace → evaluate → fix, reducing failure rates quickly.
- •Without evals/tracing, agent quality issues are invisible
- •One-command install adds instrumentation to an agentic codebase
- •Traces reveal hallucinations and wrong tool choices in real workflows
- •Using eval criteria cuts errors significantly in a short iteration loop
- 13:20 – 15:38
Avatars for UGC + internal Loom-style communication
Meng connects avatars to the rise of UGC-style content and the demand for authenticity at scale. He explains how AI avatars remove the friction of being camera-ready while still delivering a human presentation. For PMs, avatars combine with screen recording to create more engaging async updates and training.
- •UGC shift: audiences prefer human-feeling communication over corporate messaging
- •AI avatars reduce production friction (appearance, lighting, timing)
- •Tools mentioned: HeyGen (avatar/presentation), Seedance (creative filmmaking)
- •PM use case: pair avatar + screen share for async team communication
- 15:38 – 27:06
Plugins vs skills vs computer use (and the trust/security tradeoff)
Meng distinguishes between plugins (deep integrations) and skills (lightweight, user-creatable capability packs). He argues "computer use" is the most universal layer because it works anywhere a human can click, without special integrations. He also discusses security concerns (prompt injection) and why permissioning/guardrails matter for trust.
- •Plugins: productized integrations (e.g., Slack/Linear/Chrome)
- •Skills: modular add-ons anyone can create; encode domain/taste knowledge
- •Computer use: universal automation layer that can operate any UI
- •Security and trust: permission tiers and guardrails reduce risks like prompt injection
- 27:06 – 30:38
Building a project-folder system for everything (work, company, life)
Meng walks through a practical folder taxonomy under a single Projects directory, designed to keep context scoped and token costs manageable. He advocates separating content creation, customer support/issue tracking, company admin, skills, and even family/school artifacts so agents can work with the right context. The goal is a scalable, local-first operating system for AI work.
- •Create a single Projects folder and treat each project as a local context unit
- •Separate content (scripts/MD), support/issues, company admin, skills, etc.
- •Context scoping reduces token waste and confusion
- •Local organization enables reuse across Codex, Obsidian, and other tools
- 30:38 – 30:57
Permission tiers + token strategy: speed vs accuracy vs autonomy
Meng explains Codex permission settings and compute levels (low/medium/high/extra high) as levers for cost, speed, and reliability. He contrasts OpenAI’s generous usage with Claude’s tighter limits, and describes how he runs many agents in parallel on higher tiers. The key is matching permission/compute settings to the task and trust level.
- •Default permissions for low trust; full access for high autonomy workflows
- •Compute levels map to token cost, latency, and quality
- •Meng’s approach: higher-tier plan supports many concurrent agents
- •Guideline: choose settings based on task difficulty and budget
- 30:57 – 32:40
Plan Mode in action: from app idea to implementation-ready blueprint
Meng demonstrates starting with Plan Mode to avoid premature coding and to force clarity before execution. Using a QR-scanner-to-email app example, he shows how to review the plan, ask targeted questions, and refine requirements. He emphasizes that “technical” now means understanding systems and jargon, not writing code.
- •Always start with planning to approve scope and approach before building
- •Use follow-up questions to deepen unclear parts without triggering execution
- •Example: QR scan → email content + potential screenshot feature
- •Modern technical skill = understanding architecture, tradeoffs, and terminology
- 32:40 – 41:45
Slides from real data: local-file context + structured planning
Meng shows how to generate presentations by pointing Codex to local folders containing business metrics and prep docs. He reinforces that even “simple” outputs like slides benefit from planning first, and that voice input enables richer context than typing. The workflow turns scattered local artifacts into a coherent deck.
- •Codex can read local project folders to pull metrics (revenue, users, conversion)
- •Plan slides first: define structure, deliverables, and constraints
- •WhisperFlow accelerates high-context input and remembers corrected names/terms
- •Codex can output in formats like PPTX and interoperate with Keynote/PowerPoint
- 41:45 – 47:31
The screenshot shortcut that changes iteration speed (plus Codex mobile)
Meng introduces a rapid screenshot capture shortcut that feeds visual context to agents instantly, improving design iteration. He then highlights Codex Mobile, which lets you continue project chats from your phone while tasks run on your computer (which must stay on with proper settings). The combined effect is faster feedback loops anywhere.
- •Screenshots provide dense context; "image is worth a thousand words"
- •New shortcut captures the active app window quickly for agent feedback
- •Codex Mobile connects to desktop projects; work continues on the computer
- •Requires desktop to remain on; permissions/settings control behavior
- 47:31 – 54:58
Taste skills and senior-level polish: generating, critiquing, and exporting designs
They explore “taste” skills—prebuilt guidance packs that improve typography, spacing, and overall design quality. Meng explains how AI still makes layout mistakes (e.g., cramming text), and the PM’s role is rapid critique and direction. He also shows how to close the loop by exporting drafts into controlled formats like HTML, Figma (via MCP/computer use), or PPT/Keynote.
- •Taste skills are downloadable (e.g., GitHub) and improve visual output quality
- •Human role: spot overcrowding and guide iterative refinements
- •Export paths: HTML (fast/controlled), PowerPoint/Keynote, Figma via MCP/computer use
- •Prompting pattern: draft (images) → critique → recreate in target medium
- 54:58 – 1:05:31
Build your AI digital twin: avatar images → scripts → multi-format video output
Meng outlines a forward-looking workflow for creating a digital twin using existing photos, then combining it with screen recordings and scripts to produce polished videos. He specifies multi-aspect outputs (16:9, 4:3, 9:16) with hooks and calls-to-action, reflecting modern content distribution needs. He also warns about ethical lines and potential regulatory backlash around deceptive AI media.
- •Generate a digital twin with GPT Image 2.0 from curated photos
- •Pipeline: screen recording + markdown script + avatar + video assembly tool
- •Multi-format deliverables tailored to platforms (YouTube, X, TikTok/Reels)
- •Ethics: avoid deception; expect pushback on fake products/claims
- 1:05:31 – 1:12:24
PM survival in layoffs: technical PMs, founder path, and the fleet-of-agents mindset
Meng argues layoffs are disproportionately affecting non-technical PMs, and that PMs must become “technical” in the modern sense: fluent in AI tooling, concepts, and orchestration. He suggests the long-term endgame is building your own business, with AI handling much of the operational burden while you provide taste, QA, and domain expertise. He frames success as pushing beyond the rising baseline toward “11-star” quality.
- •Technical PMs survive; "technical" now = tool fluency and system understanding
- •AI reduces bureaucracy value; increases value of orchestration and judgment
- •Founder path: AI handles ops/marketing/admin; human focuses on moat and QA
- •Quality bar rises: aim beyond baseline with taste and strategic vision
- 1:12:24 – 1:14:42
Wrap-up: where to follow Meng and keep building
Aakash closes by directing listeners to Meng’s social channels and products, reinforcing the theme of persistent iteration and high prompt volume. Meng notes each product required extensive prompting and encourages people not to quit after the first attempt. The episode ends with Aakash’s standard show and bundle call-to-actions.
- •Where to find Meng online (X/LinkedIn/Instagram/TikTok) and his products
- •Expectation-setting: real outcomes take thousands of prompts and iteration
- •Podcast closing CTAs: subscribe, review, comment, and check the bundle
