Mike Krieger describes moving from step-by-step task delegation to expressing an end state and letting Fable go work on it, including porting a whole Python labs codebase to TypeScript over a weekend. He argues teams must learn to be more unreasonable in how they prompt, that Anthropic increasingly works through async 'tag' delegation and Claude Code artifacts rather than line-by-line review, and that Anthropic Labs runs on two-week persevere-or-pivot bets.
Krieger frames his own journey as both a model shift and a role shift: after roughly two years as Chief Product Officer he kept spending weekends building with the models, felt mounting FOMO, and decided to make a change right as internal snapshots of what became Mythos and Fable started appearing. He notes it's now a broader trend, with CTOs from other companies joining Anthropic as individual contributors.
The behavioral change he emphasizes is moving away from taking an idea, breaking it down in his head the way he'd normally do engineering, and iterating step by step — toward describing the end goal, letting the model go work on it, and then discussing the trade-offs it surfaces. He jokes that Fable, re-enabled only a couple of days earlier, is 'way way smarter' than him, sometimes returning finished work with trade-offs he has to ask it to explain more simply.
Riffing on Tariq's 'be unreasonable' framing from the day before, Krieger argues the industry has to teach people — including non-technical colleagues — to ask for far more. He blames first-generation AI products for putting models 'too much in a box,' constraining tools and degrees of freedom so the model could write code but not run it or introspect its environment. He cites co-work giving knowledge workers a VM that can write bash as the kind of unconstraining that lets a model remediate issues, like writing a script when a built-in PDF parser fails.
His flagship example: a labs project he'd written in Python (a nod to Instagram's Python roots). Realizing Claude Code had a better deployment story with bun, he asked it to port a few-hundred-thousand-line codebase to TypeScript. He built a dynamic workflow that ran over the weekend — porting, verifying, double-checking, and churning through the code — and returned Monday to a completed, deployable port. On whether you could port a product like Instagram to PHP, he points to Instagram's own history: adopting types in Python 3 and building 'monkey type' to capture real production types, suggesting conversions can lean heavily on production data and segmented tests, with the hard part being finding an incremental boundary rather than boiling the ocean.
Krieger says he's excited that 'tag' is finally public because it reflects how Anthropic has worked for a while — something he previously struggled to describe on stage. While Claude Code is used for interactive, high-bandwidth iteration, most usage is delegating asynchronously via tag. He compares its multiplayer nature to Midjourney on Discord: seeing others tag Claude to own part of a codebase, monitor a feedback channel, and proactively take on tasks made him realize he'd been under-utilizing it as a 'glorified Claude Code in Slack.' The advanced version treats Claude as a teammate that holds context, has memory, and can be proactive.
On review bottlenecks, he says they're constrained less by carving out review time and more by humans' ability to fully conceptualize what's being built. That's part of why they shipped Claude Code artifacts: instead of sending a 2,000-line PR, you share an artifact explaining the intention and trade-offs, treating code as ultimately verifiable while discussion focuses on intent and production measurement. He admits he doesn't review every line — he talks to Claude about the code, asking it to investigate his questions — a human-driven, Claude-powered review, with cosmetic changes handled 'fix forward.'
Asked to 'draw the org chart,' Krieger describes Labs as running on two-week reviews where every project goes up for 'persevere or pivot,' with projects shut down basically every cycle so that winding down feels normal rather than a personal failure. Because aligning the org chart to individual projects would mean reorging every two weeks, teams instead form as 'bets' that draw people from product, the engineering team, and himself when he's especially interested. A bet has a lead or directly responsible individual who usually doesn't manage the others — a deliberate break from prior norms that keeps the group flexible enough to disband easily.
He argues the engineering-manager discipline's death has been exaggerated: coaching, interpersonal work, and personal development still matter, with managers focused on making sure each individual is on the work they're most excited about. Structure solidifies only once a product gains traction — he cites Claude design, which started ad hoc, got a big second release in June, and now has a dedicated hired team. He sees Claude design's future in better interoperability between surfaces and in the blurring line between a design and a persistent, shareable app or artifact.
On why anyone should still start a company rather than join Anthropic, Krieger says a main reason he joined was that better models unlock a next generation of startups — not by solving ideation or taste, but by making experimentation far simpler and faster. Using Instagram versus a hypothetical 'very googly' Google Photos as an analogy, he argues platforms leave enormous room to be laser-obsessed with a vertical or group of users in a way no lab will match. Writing code, he stresses, was never the limiting part — space and user understanding are — so four or five people obsessed with a problem still move faster than anyone else.
He touches on vertical AI (noting Chris Lovejoy joining Anthropic's healthcare efforts) and finance, where the art is balancing verifiability, audit logging, and data provenance against the flexibility to build just-in-time analyses and dashboards. He closes on mental health and burnout in a 996 industry, advising people to actually carve out time off ('there's no job that is so important that you can't be offline for a couple of days'), to keep perspective ('you're never as good as your best game and never as bad as your worst'), and — echoing his coach — to verbalize emotions openly so teammates can too.
we have to teach people to be more unreasonable in their usage
just port this entire Python codebase to TypeScript. Get it working, get it deployable in, you know, a weekend.
The more advanced version is really trying to start think of it as a teammate that is actually sort of holds context, has memory and can be proactive.
there's no job that is so important that you can't be offline for a couple of days.
| Time | Topic |
|---|---|
| 00:00 | Intro: welcoming Mike Krieger, Fable timed for the conference |
| 00:32 | Model shift and role shift; delegating vs. describing the goal |
| 02:35 | 'Be unreasonable': unconstraining Claude and its tools |
| 04:07 | Porting a Python labs codebase to TypeScript over a weekend |
| 08:13 | How Anthropic works: async 'tag' delegation and Claude as teammate |
| 09:44 | Review bottlenecks and Claude Code artifacts |
| 11:17 | Anthropic Labs org: persevere-or-pivot bets |
| 14:52 | What would you delete in Claude? Product complexity |
| 16:55 | Why start a startup; vertical AI, finance, and burnout |