Product thinking lessons from Instagram, Artifact, and Anthropic

Executive overview

Most product failures stem not from bad technology but from solving the wrong problem, or solving the right problem too late in the user journey. Mike Krieger traces this lesson through three companies: Instagram's pivot from a cluttered location app to a focused photo-sharing product, Artifact's over-reliance on personalization before earning trust, and Anthropic's work making AI useful to first-time users.

The framework is simple: identify the real problem, prototype early, strip away distractions, and design for the user on day one — not the user after months of engagement.

Great products are built by stripping away everything that isn't working and going deeper on what is.

Finding the right problem to solve

• Design thinking at Stanford: identify problems, research solutions, validate what you built.
• Prototyping early beats six months of heads-down work followed by a painful reveal.
• Burbn (pre-Instagram) had too many features; the one thing working was photo sharing — everything else got cut.
• The pivot to Instagram came from peeling back what wasn't working, not from adding more.
• Emerging technology creates product opportunities when you find the right wrapper to make it usable for many people.

Reading the market moment

• Look for early adopters already doing something interesting in a space — that signals latent demand.
• Track the rate of change: if the technology improves two or three more steps, what becomes possible?
• iPhone camera quality jumped visibly from 3G to 3GS to 4; that trajectory told you where to build.
• Same logic applies to LLMs: Claude 3 to Claude 3.5 was a large leap, and the next leaps are coming.
• Build a product useful today that is also positioned to catch the next wave.

What killed Artifact — and what it taught

• Artifact launched after two years in private beta — too long; team was exhausted before the product even reached the market.
• The underlying algorithms were strong, but the mobile web destinations were ad-heavy and poorly formatted.
• Personalization only shone after heavy usage; new users saw no difference from Apple News and left.
• The lesson: users will not adopt a product just because the technology is intelligent — it must solve the problem from the first interaction.
• Don't bet so hard on the long-term experience that you neglect the day-one experience.

Applying Artifact's lessons at Anthropic

• Claude's onboarding asks questions upfront to personalize from conversation one, not after months of usage.
• Lower the barrier for users who have never thought about how AI can help them — give concrete, useful starting points.
• Being transparent about model limitations (what it's good at, where it makes mistakes) sets the right mental model early.
• When Claude answers medical questions, flagging "I'm not a doctor" is a design choice, not just legal cover.

Building the right team

• Seek people who are talented but low-ego — a rare and difficult combination.
• Generalists who cross disciplines (back-end engineers who build iOS, designers who code) keep the team fluid.
• People must care about the product they're building; passion drives the close attention to detail that produces breakthroughs.
• Great product ideas come from people close to the work, not from strategy meetings.
• Instagram's team culture — hackathons, experimentation, moving fast — survived the Facebook acquisition because Facebook actively preserved it.

Knowing when to shut down

• Make a concrete list of ideas you'd regret not trying before closing the company.
• Prioritise three big bets, run them, then reassess whether the trajectory has changed.
• Set a date or a defined set of projects as the decision boundary — vague timelines lead to years of drift.
• Investors often want founders to call it; the obligation to keep going is frequently imagined, not real.
• Shutting down is not a tragedy; it is a normal, expected outcome that is baked into the model.

Building AI products alongside evolving models

• At Anthropic, the product and the model are both moving targets — models can change up to a week before launch.
• Evals measure math and coding benchmarks easily; measuring "vibes" or personality requires listening to Claude with Claude.
• Hundreds of internal conversations daily track how Claude's personality evolves during training.
• Thumbs-up/thumbs-down feedback with written explanations surfaces aggregate themes: too verbose, backs down too quickly, changes its mind when it shouldn't.
• Feedback from each model cycle directly shapes the next training run.

Where AI products are heading

• Models need memory and contextual understanding that spans many conversations, not just a single session.
• The next challenge is teaching models when to be proactive versus when to stay quiet — natural collaboration cues.
• Long-running background agents represent a shift from chat-box interaction to persistent, autonomous assistance.
• AI coaching — ongoing, personalised, across time — could make the value of an executive coach available to everyone.
• The app layer will consolidate as it did with mobile and social: an explosion of products, then natural selection toward what works.