Silicon Valley GirlAI-First Playbook: Do a Team's Work With AI (2026) | Peter Yang
CHAPTERS
- 0:00 – 0:42
Hook: AI feels like cheating—and the risk of becoming dependent
Marina and Peter open with the paradox of AI-enabled productivity: it can feel like “cheating,” yet it also creates a new kind of dependence. Peter previews a personal fear—losing motivation and cognitive stamina when AI isn’t available.
- •AI makes solo work feel unfairly efficient
- •Peter’s concern about getting “dumber and lazier” with constant AI assistance
- •Framing the episode around practical, real workflows (not hype)
- 0:42 – 1:41
Self-improving “skills”: turning conversations into better instructions
Peter explains what he means by AI “skills” (simple instruction files) and how he makes them self-improving. Instead of endlessly re-prompting, he asks the AI to revise the skill based on what went wrong in the back-and-forth.
- •A “skill” is essentially an instruction text file
- •After using a skill, ask AI to update it to reduce future iterations
- •Human review is required before accepting changes
- •Self-improvement is accessible to everyday users, not just researchers
- 1:41 – 2:52
The 1-day build sprint that replaced meetings with an AI workflow system
Peter describes the turning point: taking one day off meetings to dump his workflows into Codex. He outlines the major areas he automated across his creator business and how voice-based brain dumps accelerate setup.
- •Creators repeat the same work across many formats (newsletter → descriptions → social)
- •Peter uses voice tools (e.g., Whisper Flow) to capture workflows quickly
- •Key systems: podcast prep/post, newsletter editing, cross-platform posting
- •He also built an “advisor” skill connected to business context
- 2:52 – 4:09
Cross-platform posting system: Codex scheduling, formatting, and internal APIs
Peter walks through how he posts to multiple platforms at once using Codex, even when official APIs don’t exist. The system handles platform-specific nuances (like tagging) and can operate via browser-use or discovered internal endpoints.
- •Uses cron-style briefings plus direct posting via Codex
- •For platforms without APIs (e.g., Substack Notes), Codex can use browser automation or sniff internal APIs
- •Single workflow publishes to X, LinkedIn, Threads, and Substack Notes
- •Platform-specific rules: tagging on X vs limitations on LinkedIn
- 4:09 – 4:59
Human-in-the-loop content: voice brain dumps, viral examples, and the “last 10%”
Peter explains how he drafts: he speaks rough thoughts, Codex turns them into polished posts using his prior high-performing examples, and he approves before publishing. He emphasizes that the final 10% needs human taste to avoid “AI slop.”
- •He drafts from his own ideas rather than auto-generating generic posts
- •Codex cleans and structures voice brain dumps using examples of prior viral posts
- •Approval step before cross-posting
- •Quality principle: humans must add the last 10% to avoid slop
- 4:59 – 6:28
Newsletter workflow: long-form dictation, context limits, and editing via feedback
Peter shares how newsletter posts differ from social posts—often starting with a longer, more personal voice dump. He discusses practical constraints (context windows) and why giving feedback to AI can be harder than editing directly, yet still faster overall.
- •Uses 10-minute recordings (e.g., Super Whisper) then paste into Codex
- •Avoids live dictation into Codex to prevent context/processing issues
- •Iterative refinement: he provides feedback until it’s close
- •Manual polish still required at the end
- 6:28 – 7:51
Weekly strategic brief: revenue, content performance, and competitor outliers
Peter describes an automated weekly report that compiles revenue, recent content performance, and comparative channel analysis. For sites without APIs, he uses browser automation to collect metrics that would otherwise take significant manual time.
- •Weekly briefing includes: money made, past 3 days performance, competitor/channel outliers
- •Runs across YouTube and Substack
- •Browser automation fills gaps when no API exists
- •Saves roughly an hour of manual work per week
- 7:51 – 9:13
Editing support: using transcripts to find cuts, timestamps, and “spicy quotes”
Peter explains how AI helps his video workflow even though he still uses a human editor. By feeding transcripts into Codex, he gets suggested cuts, timestamps for awkward moments, and highlighted quotes for intros and hooks.
- •Keeps a human editor, but AI accelerates decision-making
- •Transcript → suggested cuts and timestamps
- •AI highlights “spicy quotes” for intros and packaging
- •Overall production speed is dramatically faster than a year ago
- 9:13 – 10:09
From creator to builder: deciding what to build and how long it takes
Peter discusses when he chooses to build tools versus accept the status quo, focusing on solving personal pain points first. He gives examples like a fitness tracker and a scam-detection extension, plus a realistic view of “80% done” vs polished quality.
- •Build tools to solve your own recurring problems first
- •Examples: personal fitness tracking + automated health emails; scam-hiding Gmail extension
- •AI can get to 80% in hours with a clear plan
- •Polish, testing, and quality can still take days
- 10:09 – 12:16
The 5 layers of AI adoption: from Q&A to fully integrated personal agents
Peter lays out a five-level framework for AI adoption, moving from simple question-answering to deep automation. The key transition is from chat outputs you copy/paste into real integrations that execute work across your apps.
- •Layer 1: everyday answers in ChatGPT/Claude
- •Layer 2: daily work via projects—still lots of copy/paste
- •Layer 3: prototyping products with modern AI dev tools
- •Layer 4: building apps (often personal apps)
- •Layer 5: integrated agent workflows with automations + integrations
- 12:16 – 15:18
Practical upgrade to Level 5: switch to Codex/Claude Code + create a “Personal OS” folder
Peter offers a concrete path from Level 1–2 toward Level 5: stop treating AI like a chat tab and start using agentic coding environments with skills and integrations. He explains how he structures workflows in a “personal OS” and chains skills together for repeatable output.
- •Step 1: move from ChatGPT/Claude to Codex or Claude Code
- •Codex/Claude Code aren’t just for coding—often used for operational tasks
- •Create a “personal OS” folder to store skills and workflows
- •Chain skills: editing + fresh research + anti-slop + social repurposing
- 15:18 – 17:59
Inside a skill: personal advisor + Learnings.md as durable memory
Peter demonstrates a simple but high-leverage “personal advisor” skill that references his business plan and principles. He uses a Learnings.md file to store condensed takeaways from conversations, managing context window limits while keeping insights readable and explicit.
- •A powerful skill can be “just a text file” with clear role + tone
- •Leverages external docs (Google Doc plan/principles) for consistent advice
- •Leverages Learnings.md to store small, curated memory over time
- •Context management: keep saved learnings short and intentional
- 17:59 – 21:48
Building better skills without creating bloat: triggers, a skill-builder, and a skill-editor
Marina and Peter dig into the hardest part: not writing the file, but designing triggers and nuance. Peter explains he uses AI to draft skills, then uses a “skill editor” approach to remove duplicates and keep each skill short enough to audit.
- •Hard problem: deciding triggers, boundaries, and what belongs in a skill
- •Peter uses AI to create skills but worries about uncontrolled bloat
- •He keeps skills to ~one page so he can actually review them
- •Uses a “skill editor” to remove duplicates and overly broad instructions
- 21:48 – 24:36
AI chief of staff: wiring inputs (Slack/meetings) to outputs (accountability and summaries)
They explore what it would take to build an AI chief of staff that learns decision-making style and monitors execution. Peter frames it as an inputs/outputs problem and recommends starting with integrations across comms and meeting transcripts.
- •Goal: AI that observes decision-making in Telegram/Zoom and reinforces accountability
- •Example: weekly system that flags team blockers and undelivered commitments
- •Define inputs (Slack, meeting transcripts) and desired outputs (summaries, nudges, ownership)
- •Set up integrations first (e.g., Slack + meeting recorder transcripts)
- 24:36 – 25:40
The solopreneur era: what work looks like when AI handles the drudgery
Peter reflects on shifting from big-company collaboration to high-leverage solo execution. He expects to keep working hard, but with more time spent on energizing tasks and with AI acting as a consistent partner and pair programmer.
- •AI enables “solopreneur” scale without a full-time team
- •Frees time from repetitive operations into creative/building work
- •AI as a partner that knows your context and preferences
- •Aim is not less work—more enjoyable work
- 25:40 – 26:27
Peter’s honest fear: cognitive atrophy and losing motivation without AI
Peter shares his primary anxiety about AI: reliance that reduces independent effort, especially when disconnected. Marina relates this to broader societal trends toward outsourcing chores and thinking to tools.
- •Fear: dependence makes him less able/willing to work without AI access
- •Offline moments (e.g., flights) expose reliance on the “AI partner”
- •Parallel with other convenience tools (cleaners, robot vacuums)
- •Tradeoff: convenience can free time, but may weaken fundamentals
- 26:27 – 28:04
Raising kids in the AI era: protect fundamentals, learn by building, embrace failure
They discuss how children growing up with AI may skip foundational skill-building unless guided intentionally. Peter proposes a radical approach: push kids toward building and entrepreneurship so they develop resilience, judgment, and real-world learning through failure.
- •Kids can use advanced AI immediately, unlike today’s adults who learned pre-AI fundamentals
- •Need to intentionally teach critical thinking and basics
- •Peter’s radical solution: encourage building and starting small businesses/projects
- •Emphasis on learning through mistakes, not just classroom instruction
- 28:04 – 29:33
One-week action plan: stop consuming AI news—build a workflow system and iterate
Peter closes with a practical plan: get access to strong tools, start using Codex, and build workflows that save time immediately. He warns that results require upfront setup and patience, but the system compounds because AI improves and your skills/memory accumulate.
- •Buy access (e.g., $20 plan), download Codex, start building immediately
- •Shift from consumption (news/courses) to output (workflows/skills)
- •Expect iteration—AI won’t one-shot perfect results
- •Compounding advantage: skills + memory improve over time; AI doesn’t “leave” like an employee
