Aakash GuptaGemini Gems Masterclass with the Creator at Google: 3 Gems You Must Build
CHAPTERS
- 0:00 – 3:34
Gemini Gems explained: custom Gemini with built-in context
Lisa Huang defines Gemini Gems as customized versions of Gemini designed for specific use cases. She explains the core problem they solve—LLMs lacking persistent context—and why Gems reduce repetitive prompting by bundling instructions, tone, and knowledge.
- •Gems are purpose-built variants of Gemini for particular workflows
- •Primary pain point: users repeatedly re-enter role, strategy, style, and product history
- •Analogy: general contractor (base LLM) vs master craftsman (Gem)
- •Gems preserve personalized tone and context for faster, better outputs
- 3:34 – 5:26
Three must-have Gems for product managers
Lisa outlines the three highest-leverage Gems PMs should build to cover daily communication, strategic thinking, and continuous user understanding. Each Gem is tied to a core PM responsibility and becomes a reusable productivity multiplier.
- •Writing clone to draft emails/PRDs/updates in your voice
- •Product strategy advisor as an embedded thought partner using company context
- •User research synthesizer to digest interviews, surveys, and support tickets
- •Each Gem becomes a specialized assistant aligned to a PM’s workflow
- 5:26 – 6:04
How to create a Gem: instructions, knowledge, and iteration
The conversation turns into a practical setup framework for building effective Gems. Lisa emphasizes that good results come from specificity, adding the right reference materials, and treating the Gem like a product that improves over time.
- •Step 1: write clear, detailed instructions (avoid vague prompts)
- •Step 2: add knowledge by uploading key context documents
- •Step 3: test and iterate to refine outputs
- •Specialization beats one ‘do-everything’ Gem for most PM tasks
- 6:04 – 8:28
Live demo: building a product strategy advisor Gem
Lisa walks through the Gemini UI to create a new Gem, paste structured instructions, and upload example strategy/competitive/roadmap files. She then tests it with a prompt to generate product ideas and explains how to tweak until satisfied.
- •Navigate to Gems → create new Gem → name and define its role
- •Use highly structured instructions (responsibilities, approach, constraints)
- •Upload strategy artifacts (competitive teardown, roadmap, etc.)
- •Test in real time, then refine instructions/knowledge before saving
- 8:28 – 9:01
Sharing Gems with teams + common mistakes PMs make
Lisa confirms Gems can be shared across a team to amplify productivity, especially when context is common. She then lists the biggest pitfalls—weak instructions, missing context, lack of specialization, and failure to iterate—and clarifies how Gems rely on their configured knowledge rather than “learning” implicitly over time.
- •Gems can be shared to scale best-practice context across a function
- •Mistake: vague instructions; fix with examples and specificity
- •Mistake: not uploading the differentiating context that personalizes outputs
- •Prefer multiple specialized Gems; iterate like any internal product
- •Gems reference their instructions/files; update them as reality changes
- 9:01 – 11:33
Why Gemini Gems were built: discovery, personas, and reuse
Lisa shares the 2023 origin story: users struggled to discover what LLMs could do and kept re-prompting for context. The team pursued customizable personas and shareable setups so people could save useful configurations and learn from each other.
- •Goal: help users discover capabilities and reduce repeated setup work
- •Persona-based framing (doctor, writing tutor, etc.) improves mental models
- •Customization enables saving context for repeatable workflows
- •Sharing supports organic discovery inside teams and industries
- 11:33 – 16:56
Gemini Gems vs ChatGPT custom GPTs: different product bets
They discuss how OpenAI’s custom GPTs and GPT Store pushed an app-store ecosystem narrative, which influenced urgency and execution. Lisa explains Google’s differing view: instructions are easily copied and monetization was unclear, so Gems focused on personal and team productivity rather than a marketplace.
- •ChatGPT custom GPTs triggered faster timelines (“go faster”)
- •OpenAI framed GPTs as an app ecosystem with a store and monetization
- •Google bet on productivity amplification over a proprietary marketplace
- •Concern: instructions/knowledge could be copied or extracted in practice
- 16:56 – 21:16
PM career philosophy: curiosity, PM archetypes, and company cultures
The episode shifts to career lessons: Lisa attributes her path to curiosity rather than a master plan and advises PMs to tailor careers to their strengths. She contrasts Apple’s craftsmanship bar, Meta’s data + execution culture, and Google’s technical expectations for PMs.
- •Career principle: follow curiosity and learning over rigid planning
- •PM roles vary widely; identify your archetype (0→1, craft, platform, etc.)
- •Apple: extreme product bar and craft obsession
- •Meta: experimentation, metrics rigor, and fast execution
- •Google: deeply technical PM expectations and close eng partnership
- 21:16 – 23:07
Hiring AI PMs: fundamentals plus grit and growth mindset
Lisa explains what she screens for when hiring AI PMs: core PM competencies are table stakes, and differentiators are resilience, ownership, and the ability to push through ambiguity. She describes how these traits show up clearly in how candidates discuss past work.
- •Baseline requirements: strategy, metrics, execution, cross-functional influence
- •Differentiators: grit, growth mindset, and high ownership
- •Signal: candidates who push beyond expectations and learn fast
- •Fit matters in fast-changing environments where problem-solving dominates
- 23:07 – 25:27
Building Meta’s AI assistant for Ray-Ban smart glasses: constraints and bets
Lisa recounts the early development of the Meta assistant for Ray-Ban Stories (2019–2021), including skepticism that AI would be the key feature. She highlights the complexity of partnering with an eyewear giant, privacy concerns of face-mounted cameras, and hardware constraints like size, comfort, and fashion.
- •Vision: assistant as the core interface layer for smart glasses (not just voice input)
- •Funding and alignment required amid internal disagreement
- •Major partnership dynamics with EssilorLuxottica and executive involvement
- •Key challenges: privacy/bystanders, interaction design, and extreme hardware constraints
- 25:27 – 27:33
Cloud vs on-device AI for AR + how to build in a fast-moving space
Lisa predicts a shift toward on-device AI in AR for privacy, comfort, and technical reasons, while acknowledging cloud is common today. She advises PMs to avoid tech-first thinking: deeply understand both user needs and technology, then ship, test, and iterate quickly as assumptions change.
- •On-device trend driven by privacy expectations and practical deployment benefits
- •AR constraints shape which AI features are feasible and acceptable
- •Best products sit at the intersection of real user value and tech capability
- •Execution principle: ship fast, learn fast, because the landscape changes monthly
- 27:33 – 30:07
JAX at Xero: a financial ‘super agent’ built on workflows and deep data
Lisa introduces Xero and describes JAX as an agent umbrella across small-business financial jobs-to-be-done. The goal is to automate manual workflows while using transaction-level data to deliver insights that help businesses survive and grow.
- •Xero context: accounting/payments/payroll platform serving ~4M businesses
- •JAX maps end-to-end financial workflows to identify automation opportunities
- •Agents aim to offload repetitive work while improving decision-making
- •Leverages invoice/bill/payroll transaction data for personalization and insights
- 30:07 – 32:21
Agent reliability in finance: accuracy, hybrid systems, and domain expertise
Lisa explains why accuracy is non-negotiable in finance and why generic LLMs struggle with math/accounting out of the box. Xero addresses this with domain-specific workflow understanding, rich internal data, hybrid architectures (LLMs + deterministic code), and robust evaluation with expert annotators.
- •Finance requires decimal-level correctness; reliability is a differentiator
- •LLMs alone are weak at accounting/math/tax; domain knowledge is essential
- •Hybrid approach: LLMs for reasoning + programmatic controls for precision
- •Quality flywheels: expert finance annotators, measurement systems, and iteration
- 32:21 – 37:19
How to measure an AI agent: quality → engagement → business impact
Lisa presents a three-layer measurement stack: baseline quality first, then standard product adoption metrics, then monetization/business outcomes. She discusses human evaluation, scalable LLM-judge style evals, and the need for use-case-specific criteria that evolve over time.
- •Layer 1: baseline quality—does it do the job correctly and safely?
- •Evals use humans plus scalable automated/LLM-judge approaches
- •Layer 2: engagement—adoption, MAU/WAU/DAU, retention, CSAT, feedback loops
- •Layer 3: business impact—revenue attribution and broader commercial metrics
- 37:19 – 52:10
The future of PM: agents everywhere, PMs as builders, and breaking into AI PM
Lisa argues AI won’t replace PMs because product judgment remains essential, but teams and PM:eng ratios will compress as productivity rises. She advises PMs to become hybrid builder-designers, use AI tools personally even without company support, and stand out via real side-project work, networking, and interview practice for big-tech processes.
- •AI changes execution tasks, but product judgment/taste remains the PM’s value
- •Org structures will compress; fewer people can ship more with AI tools
- •PMs expected to prototype/build (design + code) rather than only coordinate
- •No-AI-at-work isn’t an excuse: use off-the-shelf tools, build on your own
- •Stand out by doing real work (research, prototypes) and showing passion/grit
- •Roadmap: get AI reps at work or via side projects, build network, drill interviews
