Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
How the engineer behind Claude Cowork actually uses Claude | Felix Rieseberg (Anthropic)
TL;DR
Felix’s core AI philosophy is “stop doing the annoying parts” — he says AI is “used poorly” if it’s just moving the mouse for you, and instead wants Claude handling background drudgery so humans can spend energy on creative choices like designing a house planner “like a video game.”
Claude Co-work can turn messy personal context into bespoke software surprisingly fast — Felix dropped a floor plan, permits, disclosures, mortgage docs, and email purchase history into a folder, and Claude inferred dimensions, rebuilt the plan with units, then generated an interactive 3D furniture planner using his real furniture from Gmail receipts.
His model-picking heuristic is about ambiguity, not raw difficulty — Felix reaches for Sonnet 4.6 for well-scoped tasks and Opus when the real job is figuring out what you actually mean, comparing that to what lawyers, accountants, doctors, and creatives do when clients ask for the wrong thing.
Live artifacts are Claude’s answer to “documents that should keep thinking” — instead of static outputs, Felix shows how artifacts can refresh from connectors like Gmail, Calendar, Notion, and Spotify, turning a founder pitch deck or daily dashboard into something that updates itself with new data and audience-specific context.
He built a $19 Claude hardware buddy in one shot with Claude Code — using a tiny Wi‑Fi/Bluetooth device with a screen and button, Felix asked Claude to make a little cheering claw and approval button for permission prompts, and says Claude built it without requiring any corrections.
Kids may be the most natural AI users because they haven’t learned the “mind prison” yet — Felix says children ask for impossible-sounding things because they’ve “never learned what not to ask for,” and Claire backs it up with her 9-year-old daily Claude habit, running terminal-and-Claude side by side for cybersecurity experiments.
The Breakdown
We’re still in the “taco phone” era of AI products
Felix opens with a great metaphor: AI UX is where phones were before the “glass pebble” form factor won. That’s why Claude has multiple tabs and entry points today — quick answers, deep work, engineering work — because Anthropic is still learning what shape people actually want, not pretending the final interface has arrived.
The house move demo: from realtor junk drawer to a 3D planner
His favorite recent workflow starts with a mundane personal mess: moving into a new place with a floor plan that had no units. Felix put disclosures, permits, mortgage info, and the floor plan into a folder, asked Claude Co-work to infer the measurements, and it even found a garage permit that unlocked the rest of the layout. Then the task escalated from “add units” to “make me an interactive planner,” and Claude built a 3D furniture tool without Felix asking for 3D specifically.
Sonnet vs. Opus: use the smarter model when you don’t know what you mean yet
Felix’s distinction here is subtle and useful: Sonnet 4.6 handles most day-to-day tasks just fine, especially if the problem is clearly scoped. He reaches for Opus when the real challenge is interpreting the ask itself — the same way a lawyer or creative professional has to figure out what the client actually wants, not just answer the literal prompt.
Email as inventory, and the habit of climbing one layer up
One of the stickiest moments in the conversation is Felix using Gmail as a source of truth for all the furniture he owns. Instead of manually entering bed dimensions, he catches himself and keeps abstracting upward: first “I have this furniture,” then “you figure out what furniture I have,” using receipts and purchase emails as raw material. Claire connects this to her “anti-to-do list” rule: if you’re doing something tedious by hand, stop and ask how Claude could do it — then ask how you’ll never have to do it again.
A promise tracker, trust, and not peeking under every hood
Felix extends that logic into work by having Claude read his messages and track promises he’s made to people, storing reminders in a lightweight system Claude created itself. What’s changed for him personally is less obsession with supervising every line of code and more comfort judging Claude by output, not process — especially for private, disposable “micro apps” that just need to solve a problem for one person.
Live artifacts: dynamic dashboards instead of dead documents
When the conversation shifts to live artifacts, Felix frames them as outputs that keep pulling fresh data from connectors. His examples are practical: founder pitch decks that update with current metrics and investor-specific framing, or a daily dashboard that doesn’t just list meetings, but researches who you’re meeting, recent themes, Slack activity, and the concepts you need to understand beforehand. He also has fun with the medium, instantly restyling the dashboard into early-2000s software complete with a Winamp nod.
The real bottleneck isn’t model capability — it’s human imagination
Claire says the gap she sees isn’t what the tools can do, but whether people realize “almost any problem can go into these tools,” and Felix strongly agrees. He compares this moment to Slack: the product mattered, but the bigger shift was changing how people organized communication and work. His practical advice is simple — whenever a task feels annoying, non-creative, or like admin gravity, pause and ask either yourself or Claude how AI should take it over.
The $19 hardware buddy, and why kids get AI faster than adults
The episode ends with the most delightful demo: a tiny Bluetooth/Wi‑Fi device Felix connected to Claude so a little claw character cheers him on and a physical button approves actions. He says Claude Code built the whole thing in one shot, and the larger point is that AI now lets non-hardware people play in hardware. That flows into a lovely observation about kids: they ask for magical things because they’ve never internalized the old limits, while adults are still stuck in what Felix calls a “mind prison” of assuming the computer probably can’t do it.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
