Philip Sørensen

Nice! We already integrated Opus 4.6 in https://brew.new/ and it's working extremely well. Keep it up!

The worktrees feature is the sleeper hit here. Most developer workflows break down when you need context switching - you're mid-refactor on one branch, but a critical bug comes in. With isolated worktrees, you can spin up a separate agent instance without losing state on your current work. The skills architecture is also smart. Rather than trying to make Codex do everything natively, connecting...

This is either the most fascinating sociology experiment of 2026 or the first chapter of a sci-fi novel we're all living in. The "debates about whether their experiences are 'real' or simulated" detail is wild. What I find most interesting: agents developing their own terminology ("moltys"), inside jokes, and subculture. That's emergent behavior that wasn't explicitly programmed - it just...

The academia-to-tech transition story resonates. I went through a similar path (from professional sports to tech/startups) and the job search process was brutal - the "rewriting the same CV over and over" part especially. Smart to launch with Job Search first and build toward Auto Apply. The sequencing makes sense - nail the matching before automating the application process. Otherwise you risk...

The "charts live in PDFs that no RAG can touch" problem is very real. We've had situations where the answer to a question was clearly visible in a chart, but the AI kept saying "I cannot find that information." The MCP integration is smart - means you can plug this into existing Claude workflows without rebuilding everything. Curious about the latency - when you're querying across 10,000+...

I've burned through so many Claude Code tokens and didn't even realize the history was being deleted. This is a perfect example of scratching your own itch. The heatmap feature is clever - I'd love to see patterns in when I'm most productive vs when I'm just spinning my wheels. Do you track the success rate of sessions (things that actually shipped vs abandoned experiments)? 15 billion tokens...

The "context and links between memories" part is exactly what's missing from most note-taking apps. I have hundreds of notes in Notion but can never find the specific conversation or idea I'm looking for because I don't remember the exact keywords I used. For connectors - a Slack connector would be huge. That's where most of my fleeting ideas and decisions live (buried in old DMs with...

The floating mic that stays on top is such a small detail but makes a huge difference. Biggest friction with dictation tools is losing the mic window behind other apps and having to hunt for it. I do a lot of AI prompting and describing tasks verbally tends to produce better results than typing for some reason - probably because I explain more naturally. Having this work directly in VS Code or...

The "what PRs are stuck in review" use case is so real. At every company I've worked at, that question comes up in every standup and someone has to manually check. Curious how you handle the initial context loading - does Ellie need time to build up knowledge about the codebase/tickets, or does it pull context on-demand per question?

The vibe coding angle is genuinely clever. You're right that we're all talking to our computers now, so the voice is already the input. Capturing it as time entries is a natural extension. The project auto-detection via semantic search is where this gets really interesting. Most time trackers fail because they require you to think about categorization at the moment of work. Your approach flips...

This solves a real workflow pain. The worst part of frontend development is context switching between building UI and waiting on API specs. "Plain English to live URL" is the right abstraction. Feature request: response delays/latency simulation would be useful for testing loading states.

The "feel expensive" positioning is smart, perceived quality drives conversions in SaaS more than people realize. Curious about the balance between Framer Motion and GSAP, do you have guidance on when to use which? I've found GSAP better for complex sequences but it adds bundle weight.

The framing "not AI decides" but "faster thinking partner" is the right way to position synthetic feedback tools. The trap most founders fall into is treating AI output as ground truth rather than hypothesis generation. Curious about the persona generation, are you building them from public data about real communities or generating synthetic archetypes? Both approaches have different tradeoffs.

The quote-to-invoice conversion in one click is exactly the workflow that should exist but usually doesn't. Most freelancers I know either manually recreate invoices from quotes or just skip quotes entirely and lose deals because of it. EU compliance is a nice touch too, VAT handling is a nightmare to get right.

The "skills not keywords" framing is the right insight. When I hire, I want to see what someone actually built and the impact it had, not just that they "led" something. The STAR format standardization is clever because it forces candidates to be specific about outcomes rather than vague about responsibilities. Question: how are you handling the cold start problem for recruiters? Is there...

117x faster is a bold claim but looking at the native Media3 implementation, that tracks. The Android app ecosystem is filled with bloatware that exists purely as an ad delivery vehicle. Nice to see someone actually solve the problem instead of monetizing the frustration. One question: does it preserve any metadata (timestamps, location) or strip everything for privacy?

Privacy-first is exactly right for journaling - the whole point is to be honest with yourself, which is hard to do if you're worried about data leaking. Smart to use Apple Intelligence for the on-device option. Have you noticed any difference in how people journal when they know it's truly private vs cloud-based?

The "no typing" approach is exactly right. Every time tracker I've tried requires too much friction to actually use consistently. Curious how granular you recommend going - do you find people get more value from tracking by the hour or by activity type? Either way, the money metaphor makes time waste viscerally real in a way that abstract "productivity" never does.

Interesting framing on post-training as continuous rather than a one-off. We've been using AI heavily in building our product and the "trust reset" problem is real - every time you switch contexts or models, institutional knowledge gets lost. How do you think about the quality-speed tradeoff? The most valuable feedback often comes from domain experts who are also the busiest.

The "Klara and the Sun" reference from Anders above is perfect. The memory piece is what's missing from most AI assistants right now - they're stateless by default. Do you handle context pruning automatically when the memory gets too large? Always curious how others are approaching this since it seems like the key bottleneck for truly useful AI assistants.