Silicon Valley GirlClaude Engineering Lead: "Stop Using Claude Like It’s 2024"
CHAPTERS
- 0:00 – 2:05
Aim higher than email summaries: ambition as the real unlock
Felix argues that using AI for small tasks like summarizing emails misses the point of “new intelligence.” Marina frames the conversation around leveling up workflows across her large creator operation and becoming more ambitious with AI-enabled execution.
- •Felix says focusing on simple tasks (like summarizing emails) doesn’t tap into AI’s real value
- •Marina describes using Claude across channel operations and launching new distribution efforts with AI support
- •Felix contrasts building traditional software products with the open-ended possibilities users find with Claude
- •Marina sets the goal of learning more advanced ways to use Claude and AI workflows
- 2:05 – 3:04
Point AI at massive data to discover what you couldn’t by hand
Felix shares an example of someone directing Claude at a large NASA dataset with the goal of finding a new planet. He uses it to illustrate augmentation: using AI to do things you wouldn’t realistically do manually, not just speeding up routine work.
- •Felix describes a story where a person pointed Claude at NASA data to search for a new planet
- •Felix notes the finding wasn’t confirmed yet and emphasizes his own non-expertise in astrophysics
- •He reframes the example as a general pattern: aim AI at large datasets to surface unnoticed insights
- •The story motivates thinking beyond typical “knowledge work” use cases
- 3:04 – 4:16
Felix’s current workday: prototyping at higher altitude
Felix explains he no longer writes code and instead runs workflows where agents brief him and Claude prototypes variations quickly. He spends much of his time critiquing prototypes, thinking about user experience, and dealing more with human factors like emotion and empathy.
- •Felix says he’s shifted to thinking at higher “altitude,” including more focus on empathy and human dynamics
- •He describes doing lots of prototyping: sleeping on ideas, generating variations with Claude, then critiquing results
- •He says AI reduces how often he needs to reach out to specialized professions, blending roles in his work
- •He frames his day as exploring ideas quickly and selecting what’s worth pursuing
- 4:16 – 5:44
What makes a good AI product: attention limits, fewer buttons, and “it clicks”
Asked how he judges what’s good, Felix says the calculus has changed because prototypes are cheap. With AI making it easy to build many versions, the scarce resource becomes human attention; good products are simple and “click” quickly for users, enabling new capabilities.
- •Felix says AI changes cost tradeoffs: you can fast-forward to prototypes rather than debating feasibility up front
- •He highlights attention as the constraint—teams can generate far more options than people can evaluate
- •He argues good products often have fewer buttons and remove unnecessary features
- •His key signal: users quickly ‘get it’ and can do things they couldn’t do before
- 5:44 – 7:13
Automation vs augmentation: the real leverage is workflow redesign
Felix distinguishes between automation (removing disliked or costly tasks) and augmentation (enabling new abilities). He says the most impressive outcomes come not from the tool’s features but from how people rethink and reflect on their workflows and definitions of “good.”
- •Felix differentiates automation (economic value, saving time/money) from augmentation (new capabilities)
- •He says harnessing AI requires rethinking outcomes and evaluation criteria
- •He emphasizes workflow change as more important than the tool itself
- •He praises people who can reflect on and redesign their own processes
- 7:13 – 12:37
Judging creative and knowledge work: evaluate the process, not just the result
Marina challenges how to define “good” in writing and creative work. Felix argues that correct-looking outputs shouldn’t pass if they’re guesses; instead, higher quality means using reputable sources and professional steps, and in writing, using prompts that mimic expert processes (counterarguments, magnification, disagreement).
- •Felix gives a “weather reporting” example: a correct guess is still a bad outcome if the process is unreliable
- •He suggests grading AI on method—location, sources, and reasoning—especially when verification matters
- •Felix describes how a strong writer uses Claude to stress-test arguments and generate counterarguments
- •He says the ‘secret sauce’ is often automating expert steps rather than requesting a single polished output
- 12:37 – 14:30
Catching confident mistakes: adversarial reviews and a separate ‘judge’
Felix offers a practical tactic for error-prone outputs: generate adversarial critiques by spawning multiple threads tasked with attacking the work. He then uses another Claude session as a “judge” to validate which critiques are meaningful, warning that a single session is often bad at evaluating itself.
- •Felix recommends adversarial reviews: multiple threads try to break the work and prove it wrong
- •He says it can be easier to find flaws than to create, so lower-intelligence reviewers may still work well
- •He explains using Claude Projects to launch parallel critique threads on a document or post
- •He describes a ‘judge’ workflow: another Claude session evaluates whether criticism is valid and actionable
- 14:30 – 17:45
When to trust AI (and when to throw more tokens at it)
Felix explains two trust modes: when you can’t easily verify, add more validation layers; when you can judge quality directly (like design taste), you can act more like a client reviewing output. Marina adds that the challenge becomes choosing which projects require deep step-by-step control versus letting go.
- •Felix says if you can’t verify quality, use more AI layers to validate rather than relying on one answer
- •He notes some tasks can be judged by feel (e.g., whether something looks good) without mastering the full process
- •Marina argues for focusing deep control on the most important projects rather than micromanaging everything
- •They connect this to the broader challenge of ‘letting go’ when using agents
- 17:45 – 22:18
Are we overcomplicating agents? Give broad access plus tools, not role-play personas
Felix says the industry may be over-rotating on “agents” and role-based character prompts, since generalist models are strong and personas can limit thinking. He advocates giving models broad access to business context and using tools with clear documentation so the model can fetch what it needs rather than ingest everything.
- •Felix argues generalist models can outperform specialized setups, and personas can be unnecessarily limiting
- •He says he often doesn’t assign roles; he uses threads/projects but not character-based identities
- •Marina asks about data slicing per agent; Felix recommends broad access and using tools to retrieve relevant data
- •He highlights improved tool calling and cautions that ‘skills’ can have a short half-life and sometimes hurt output
- 22:18 – 23:27
“I’m just a conveyor belt for Claude”: shift from manual handoff to model autonomy
Felix describes realizing he was manually ferrying crash reports to Claude and deciding Claude should access all crash data directly and choose what to fix. He frames this as a higher upfront setup cost that pays off by unlocking more value and reducing human bottlenecks.
- •Felix explains the ‘conveyor belt’ problem: humans manually selecting tasks to hand to the model
- •He describes the alternative: give Claude access to all crash data so it can prioritize and act
- •He acknowledges the upfront cost of setup but says it opens many doors after it’s in place
- •Marina connects this to giving models raw data and letting them make decisions rather than being overly prescriptive
- 23:27 – 26:07
Getting started fast: desktop app, built-in browser, and delaying connectors
Felix recommends a simple on-ramp: use the Claude desktop app and let Claude operate through a browser logged into your systems. He suggests avoiding complex connector setups until you’ve proven a workflow is valuable, then upgrading automation once the value is clear.
- •Felix’s step one: download the desktop app and use the built-in browser
- •He advises ‘forget about connectors’ initially and let Claude use your computer to validate workflows
- •He suggests adding connectors later only after you see a workflow consistently working for you
- •He reiterates he doesn’t rely on role prompts like ‘you’re an expert,’ focusing instead on tools and techniques
- 26:07 – 27:45
Don’t let one session grade itself: external signals, feedback loops, and side-by-side review
Felix says he’s less worried about trusting Claude in general and more worried about trusting a single session to judge its own work. He recommends using outside evidence (metrics) or another session as a check, and suggests running Claude’s feedback side-by-side with your own to build trust and refine principles over time.
- •Felix says the current session is often bad at judging the quality of its own output
- •He recommends outside signals when possible and another Claude session as a practical fallback
- •Marina describes a feedback-loop problem where she becomes the bottleneck and adds ‘taste’ manually
- •Felix proposes side-by-side reviewing: compare your feedback to Claude’s across rounds to learn differences
- 27:45 – 31:13
Making AI fit team workflows: capture feedback, distill principles, and integrate into collaboration tools
The conversation turns to operationalizing feedback across a team: where feedback lives (Telegram, Slack, GitHub) and how to feed it back into the model. Felix describes a process of aggregating historical feedback, condensing it into rules, and optionally uncovering implicit principles, and he recommends using Claude in Slack for multiplayer collaboration.
- •Marina explains feedback is in Telegram and tooling is still manual for pulling it into Claude
- •Felix says access to tools reduces the need for humans to assemble feedback manually
- •He outlines a three-step approach: gather feedback, condense into principles, then probe for implicit beliefs
- •He recommends using Claude inside Slack for a shared, multiplayer workflow where teams can talk to Claude together
- 31:13 – 37:34
What stays valuable when building becomes cheap: empathy, product sense, and replacing “garbage software”
Marina asks what becomes valuable when software can be built quickly by small teams. Felix argues the fundamentals haven’t changed: empathy for users and understanding real needs still dominate, while bad, clunky software will be most vulnerable as people can build alternatives or demand better experiences.
- •Felix says even if execution is cheaper, building a valuable product still hinges on empathy for the user
- •He predicts ‘bad software’ used only because there’s no alternative will disappear
- •He argues great products solve needs well, often in ways users wouldn’t have designed themselves
- •He points to untapped opportunity in clunky legacy systems (like those used in hotels and other operations)
- 37:34 – 43:25
Capability overhang and staying valuable: be ambitious, build real things, learn how companies work
Felix agrees there’s a large gap between model capability and typical usage, and urges individuals to ask for outcomes, not just tasks, until they hit a ceiling. For engineers and future workers, he emphasizes building real projects, focusing on impact, and learning how businesses operate and collaborate—not memorizing APIs.
- •Felix says there’s a ‘capability overhang’: models can do far more than people ask of them
- •He suggests pushing prompts upward toward the desired outcome, then learning where the ceiling currently is
- •For engineers, he recommends building a portfolio of real things and measuring success by user impact
- •He advises learning how organizations work—advocating ideas, collaborating cross-functionally—alongside technical ability
- 43:25 – 49:23
Raising kids in the AI age and a cautious view of prediction: products will look totally different
Felix shares what he’d teach his daughter: understanding how businesses operate, collaboration, and value creation, drawing on what he admires about internship-heavy paths. Asked about the next five years, he avoids firm predictions but says people underestimate exponential change and believes today’s AI products are still in an early, pre-breakthrough phase.
- •Felix says he’d push early learning about business operations, collaboration, and finding value
- •He highlights learning from real company experience and cross-functional navigation as differentiators
- •Marina says she teaches her children that anything is possible through passion and hard work
- •Felix says it’s hard to predict, but he expects AI products to look completely different than today’s interfaces
