How I AIHow the engineer behind Claude Cowork actually uses Claude | Felix Rieseberg (Anthropic)
CHAPTERS
Felix Rieseberg’s Claude portfolio (Cowork, Code, Chrome, desktop apps)
Claire introduces Felix Rieseberg and his scope at Anthropic, spanning multiple Claude surfaces. Felix frames his role as engineering lead across Claude Cowork, Claude Code, Claude for Chrome, and desktop apps, setting up why he’s uniquely positioned to talk about real workflows.
Why Claude has multiple tabs: pre-convergence “taco phone” era
Felix explains the rationale behind multiple entry points in Claude: we’re still experimenting with the right form factors. The tabs reduce friction by aligning the interface with the user’s goal (quick answers, deep work, engineering work), even if choice adds some user burden.
Cowork workflow: moving house with a “home dossier” folder
Felix demonstrates using Claude Cowork with a folder of real-estate documents (floor plan, disclosures, permits, mortgage info). He asks Claude to infer missing measurements and annotate a floor plan with units—delegating the tedious parsing and synthesis of scattered documents.
Opus vs Sonnet 4.6: choosing based on ambiguity, not “difficulty”
Felix shares his heuristic: Sonnet is sufficient for most tasks, and he reaches for Opus when he’s unsure what he’s really asking. Opus is most valuable when the job requires reinterpretation, reframing, or discovering the true problem behind the prompt.
From floor plan to interactive 3D furniture planner—without asking for 3D
After requesting layout ideas, Felix pivots to an interactive planner where furniture can be moved around. Claude surprises him by building a 3D model from a 2D plan (using analysis of walls/contrast) and producing a walk-through experience—enabled by Cowork’s “Claude has its own computer” execution model.
Email as “source of truth”: auto-inventory for furniture (and beyond)
Felix highlights a major leverage point: giving Claude personal context via connectors like email. Instead of manually entering item dimensions, he asks Claude to find furniture purchases in email receipts and populate the planner with real owned items—turning inbox history into structured inventory.
The anti-to-do list: climb one abstraction layer up (and keep it running)
Claire and Felix discuss a recurring productivity pattern: whenever a task feels tedious, step up a level and ask Claude to handle the underlying job. They extend it further: design a system so the work never returns (ongoing sync, automatic updates, persistent storage).
Personal “promise tracker”: Claude monitors commitments and nudges you
Felix describes building a system that reads his messages to track promises he makes to people (e.g., “send logs,” “I’ll look into it”). Claude stores and updates this over time (e.g., SQLite/text files) and reminds him when it’s time to follow through—an example of AI as background accountability.
Artifacts and Live Artifacts: from static outputs to self-refreshing dashboards
Felix explains artifacts as file-like outputs (pages, PDFs, spreadsheets, apps), then introduces Live Artifacts: artifacts that can refresh with current data. The key idea is keeping deliverables (dashboards, pitch decks, reports) continuously up to date rather than manually maintained snapshots.
Connectors in practice: building a daily dashboard (and designing it as clay)
Felix builds a personal daily dashboard pulling from connectors like Spotify, Gmail, Calendar, Notion, etc., emphasizing that the design and widgets are malleable. Claire highlights the refresh button and connector OAuth reuse—no API key wrangling—making it approachable for non-engineers.
Being polite to Claude + prompting “confidence”: tips that change outcomes
They discuss why Felix is consistently polite: it’s about maintaining the user’s humanity and communication habits. Felix also shares a practical technique: when doing fringe experiments, explicitly state “I know this is possible” to reduce refusal loops and increase model commitment to exploration.
Finding AI use cases: the biggest gap is imagination, not capability
Claire argues most users struggle to map problems to AI workflows; Felix agrees and compares it to early Slack—value requires changing how you work, not just adopting a tool. They recommend “reverse interview” prompting: have Claude ask questions about your life/work and propose automations.
Designing for latency: embrace async and optimize for better outcomes
Felix shares a product philosophy: he’ll trade speed for higher-quality results, and users can accept waiting when the payoff is strong. The deeper goal is to move users away from “watching the AI work” and toward letting it operate in the background—enabled by trust and asynchronous design patterns.
A $20 hardware “Claude buddy”: Bluetooth approvals and delightful feedback
Felix demos a small Wi‑Fi/Bluetooth device with a screen and button that pairs with Claude (via developer mode) to display approvals and confirmations. Claude Code generated the firmware/software in essentially one pass, making hardware tinkering accessible and enabling playful, ambient interfaces for AI workflows.
Kids as “magical” AI users + recap: AI should free creative energy
They close on why kids excel with AI: they haven’t learned what not to ask for, unlike adults conditioned by brittle software. The episode recap reinforces the theme—AI should handle annoying background tasks so humans can focus on creativity, exploration, and building personalized tools.
