How I AIHow the engineer behind Claude Cowork actually uses Claude | Felix Rieseberg (Anthropic)
CHAPTERS
- 0:00 – 2:40
Why AI should remove busywork (and free your creative energy)
Felix and Claire open with a philosophy: AI is most valuable when it quietly handles annoying background tasks rather than “driving the mouse.” They preview the episode’s practical theme—finding everyday workflows where Claude meaningfully reduces cognitive load.
- •AI is “used poorly” when it only automates superficial interactions
- •Goal: offload tedious work to reclaim creative time
- •Preview of Claude workflows and a $20 hardware “Claude buddy”
- •Framing: the real gap is translating problems into AI workflows
- 2:40 – 7:15
Felix’s role at Anthropic and why Claude has multiple “tabs”
Felix explains his scope—Claude Cowork, Claude Code, Chrome, and desktop apps—and why the product offers multiple entry points. He compares today’s AI UX to the pre-smartphone era: experimentation across form factors before convergence.
- •Felix leads engineering for several Claude client products
- •Multiple tabs reflect different user intents (quick answers vs deep work vs engineering)
- •Current era = experimentation with UX “form factors”
- •Long-term vision: fewer choices, more unified experience
- 7:15 – 9:51
A real Cowork workflow: turning a house folder into actionable plans
Felix shares a moving-related workflow: he collects floor plans, disclosures, permits, and other docs into a folder and asks Cowork to infer missing details (like dimensions). Claude uses that context to generate more useful versions of documents and jump-start planning.
- •Organize a single folder as the project’s source of truth
- •Ask Claude to infer units/dimensions from related documents (e.g., permits)
- •Use Cowork’s ability to work across multiple files in context
- •Workflow mindset: start with a concrete problem and let Claude expand possibilities
- 9:51 – 12:32
Opus vs Sonnet 4.6: choosing the model based on problem ambiguity
Claire presses Felix on when to use Opus versus Sonnet. Felix’s heuristic is about how well-scoped the problem is: use Opus when you need help reframing the question itself, not just executing clearly-defined steps.
- •Sonnet 4.6 is sufficient for most everyday tasks
- •Opus is best when you don’t yet know what you’re really asking
- •Model choice depends on tolerance for misinterpretation/reframing
- •Analogy: experts (lawyers/doctors/creatives) reinterpret the real problem
- 12:32 – 14:34
Claude generates an interactive 3D furniture planner (without being asked)
Felix shows how Claude went beyond “place furniture on a floor plan” and produced an interactive planner, then even built a 3D walk-through model. He highlights Cowork’s core capability: Claude can use its own virtual machine to build whatever software it needs.
- •Artifact output can be a full interactive app, not just text
- •Cowork provides Claude a virtual computer/dev environment
- •Claude inferred a 3D approach from a 2D floor plan using image analysis
- •Value is emergent: Claude proposes solutions users didn’t think to request
- 14:34 – 15:45
Email as a personal inventory database (furniture today, clothes tomorrow)
They explore a powerful pattern: treat email receipts as a “source of truth” for what you own. Felix uses connected email to automatically extract furniture purchases and dimensions; Claire notes broader uses like moving inventories and wardrobe tracking.
- •Email receipts/confirmations can be mined for structured inventory
- •Claude can extract items, dimensions, vendors, and dates from purchase history
- •Reduces manual entry when planning moves or organizing possessions
- •Generalizable idea: build personal data systems from existing exhaust
- 15:45 – 18:07
The “anti-to-do list”: keep climbing one abstraction level up
Felix and Claire describe a habit shift: whenever you’re doing tedious manual work, step back and ask for the higher-level goal. The aim is not only delegating the task to Claude once, but designing a system so you never have to do it again.
- •Notice tedious work → ask “what’s the higher-level objective?”
- •Replace manual data entry with “Claude, you figure it out”
- •Next step: make it persistent so it stays solved long-term
- •Mindset: spend your time steering and deciding, not transcribing
- 18:07 – 23:14
Promise tracking as an ongoing system (and letting go of perfectionism)
Felix shares a personal automation: Claude reads his messages to track promises he made and reminds him when they’re due, using lightweight storage like SQLite/text files. They discuss a key unlock—judging by outcome, not whether Claude’s internals are “architecturally perfect.”
- •Claude can monitor messages to track commitments and deadlines
- •Automation can store state in simple local databases/files
- •Shift from code-reviewing everything to evaluating real-world impact
- •AI use expands into domains where you’re not the expert (e.g., medical info)
- 23:14 – 26:11
From Artifacts to Live Artifacts: what they are and why they matter
Felix defines artifacts as file-like outputs (pages, PDFs, apps), then introduces Live Artifacts: artifacts that can refresh with updated data. The key benefit is reducing repetitive “update work” (like pitch decks and dashboards) by keeping outputs connected to live sources.
- •Artifacts = durable outputs (documents, apps, spreadsheets, PDFs)
- •Live Artifacts add refreshable data so outputs stay current
- •Use case: founder pitch decks and recurring reports that need updates
- •Core idea: keep one template and continuously refresh the underlying facts
- 26:11 – 28:41
Building a personal daily dashboard using connectors (OAuth, refresh, and remixing)
They demo a daily dashboard Live Artifact pulling from connectors (Spotify, Gmail, Calendar, Notion, etc.) without managing API keys. Claire highlights the refresh button and OAuth reuse; Felix emphasizes going beyond calendar summaries to richer preparation via cross-referencing context.
- •Connectors bring external data into Claude via simple auth flows
- •Live Artifact refresh button reloads and recomputes the dashboard
- •Better dashboards prep you for meetings by summarizing context and relationships
- •Artifacts are “clay”: style and widgets can be iterated quickly
- 28:41 – 42:21
Being polite to Claude—and a practical tip: assert what’s possible
Claire asks about Felix’s consistently polite tone; both argue it’s about preserving your own humanity and communication habits. Felix adds a practical prompting tactic: if you know something is possible but hard, explicitly state that to boost follow-through and reduce unproductive pushback.
- •Politeness is for the human’s mindset, not the model’s feelings
- •“I believe in you / I know it’s possible” can reduce back-and-forth
- •Tip: preempt refusals by framing as exploratory/creative work
- •Trust-building helps users stop hovering and let work run asynchronously
- 42:21 – 46:42
Designing around latency: make async work feel normal (and worth it)
They discuss product design principles for AI’s inherent wait time: users will tolerate latency if the result quality is high and the system is reliably trying in the background. Felix argues the real goal is to stop making users watch AI work and instead let it operate asynchronously while humans do higher-value thinking.
- •Latency is acceptable when outcomes are high-quality
- •Asynchronous retries (Slack/iMessage analogy) reduce perceived failure
- •Don’t make users “watch” the model; build for background execution
- •Trust > patience: confidence in completion changes user behavior
- 46:42 – 59:25
A $20 hardware Claude buddy: approvals, notifications, and kid-friendly interfaces
Felix demonstrates a tiny Wi‑Fi/Bluetooth device that pairs with Claude to signal when approvals are needed and celebrate progress—built via Claude Code with minimal hardware expertise. They close by reflecting on kids as fearless AI users, and end with prompting/feedback tactics (debugging with Claude and using thumbs-down for product improvement).
- •Low-cost hardware + Claude Code enables custom physical companions
- •Device supports approval workflows with a big physical button
- •Kids are “magical” AI users because they haven’t learned learned helplessness
- •When Claude goes off-rails: debug expectations and improve harness/prompt
- •Use thumbs up/down feedback to help improve Claude products
