Observations
Not ideas, findings.
Patterns, gaps, and friction points noticed while building with AI. Each observation is a documented finding with an implication worth tracking.
Strategy sessions generate obsolete code with no way of knowing which is which
When you're deliberating, actively researching, and pivoting strategy inside a chat on a large language model, you end up with code or CLI prompts that were generated earlier in the conversation but are no longer relevant. Further down in the same conversation you changed direction, which generated new code going a different way. The difficulty is compounded when you work the way I work: strategizing from my phone, then going to my computer to implement whatever came out of those sessions. You sometimes have to run all the code sequentially to make sure you don't miss anything, or try to manually remember which parts of the conversation are obsolete due to a mid-session pivot and which are still fresh and applicable. It would be a genuinely useful feature for builders if these models automatically detected code or prompts that no longer match the current direction of the build and put a strikethrough on them. So you don't have to manually track this yourself or waste credits and time running prompts you don't need anymore.
The Walled Garden Problem
During the process of building my personal website and fleshing out content to publish there and use on social media (for both my personal accounts and VYNS) I realized something. Most of it came from conversations with AI, and none of it could be easily exported to an agent I could run for discovery on which thoughts and ideas across the different large language models I use are worthwhile to publish as content. The only way to do it is to manually copy-paste, which is time-consuming. OpenAI has an export function and I could route that to an agent, but there would be a lot of context drift and training required to get it right, and that's only one model. This is another major limitation of every major LLM platform being a closed system. As a solo founder bootstrapping with basically zero funding, I'm always looking for practical, cost-effective ways to save time while building. For most solo founders, I imagine creative thinking and problem-solving happens inside these closed systems. I've seen a move toward more open-source models lately, but I haven't done a ton of investigative work there yet. What I do know is that the current landscape of closed models makes a layer that sits above everything and treats your conversations across every platform as a unified stream of thought awesome to imagine but currently impossible to build due to API read limitations on your own private data. For now the workaround is obvious and a little more time-consuming. I personally monitor my chats, and when something good surfaces I copy-paste it into my personal branding chat and evaluate it as content for my website or socials. Hopefully in the near future an option opens up that allows for automating some of this. It would free up time and, as the system gets optimized, make sure nothing gets missed.
Native LLM voice chat doesn't offer transcription, and it should
Native LLM platforms that have built-in voice chat functionality don't offer transcription as a built-in feature or add-on. This could be useful for power users or anyone generating specific outputs from their conversations for media and marketing. This morning I spoke with ChatGPT on my morning walk. I'm turning that into a podcast and have to use an external transcription service just to get the text for my site. A tool that hooks into the API, captures voice chat in real time, and parses it into publishable content (blog posts, podcast transcripts, build logs) would be valuable. I might eventually build this into VYNS product offerings down the road if I don't see it emerge on its own within native LLM offerings.
The hidden cost of context
Anthropic markets the 1M token context window as a capability milestone. And it is one. Being able to feed an entire codebase, a full document collection, or months of conversation history into a single API call is genuinely useful, especially when your whole value proposition depends on deep context, as mine does with VYNS. But for most of the last year, that capability came with a hidden trip wire. Once a request crossed 200K tokens, the entire call shifted into a premium pricing tier, a 2x multiplier that applied retroactively to the whole request, not just the overage. Anthropic removed that surcharge in March 2026. That's a good move. But the way it worked before and the way the change was communicated (a pricing page update, not an announcement) says something. Around the same time, they quietly adjusted how usage limits burn during peak hours. If you're building on weekday mornings Pacific time, your session capacity is used up quicker than at other times, while your weekly total stays the same on paper. There's no real-time visibility into this. A lot of users are frustrated and hoping Anthropic provides a dashboard or some real-time metric to view token burn and actual usage. What you ended up feeling was hitting a wall much earlier than the documentation implied. Anthropic has described the opacity of usage limits as a 'deliberate product decision.' Not sure why they went this route. The practical move is to build like the pricing will change again, because it will. Instrument every API call. Build cost ceilings by operation type. Never let a 1M context window become the default just because it's technically available. The capability is real. The bill is also real. Make sure you know which one you're actually using.
The most powerful AI tools are being rationed, and solo builders aren't in the first cohort
Claude Mythos leaked this week. Anthropic confirmed it's real, described it as a 'step change,' and said it's currently being tested with a small group of early access customers. That group isn't me. It probably isn't you either. This is new. A year ago, every model Anthropic shipped was effectively available to everyone with an API key on release day. The delta between what a funded enterprise customer could access and what a solo bootstrapped builder could access was close to zero. Mythos changes that. The most capable model ever built is being distributed on an invitation basis, tiered by relationship and use case, with general availability deferred while the cybersecurity implications get worked through. I don't think this is wrong. A model that can find and exploit software vulnerabilities faster than human defenders probably shouldn't be available to everyone immediately. The deliberate rollout makes sense. But I want to name what it means for the builder layer: the gap between what well-resourced teams can build and what solo founders can build just got wider, not because of money, but because of access. The most powerful reasoning and coding capabilities are going to the companies already in the room. The rest of us build with last quarter's model. One more thing: Anthropic left the announcement of a model with unprecedented cybersecurity capabilities in an unsecured, publicly searchable data store. The irony needs no elaboration.
Projects fragment across chats, and there's no native way to chain them
Everything I am building for VYNS involves multiple parallel workstreams across multiple chats. Product, infrastructure, brand, legal, GTM, marketing, build and deployment prompting, and more. Each lives in a different chat thread, often across different AI tooling. The problem is that there's no native mechanism to hand off state between sessions. You can't chain outputs from one conversation into another without significant manual effort and time to get the context right. I currently manage this with a combination of hand-written notes, copy-paste, and session summaries that I paste back in to start again. That friction is real and I estimate it is roughly a 20-30% tax on reconstruction time. It compounds daily.
AI as technical co-founder is real, but the context window is the bottleneck
I use Claude as my main technical co-founder for the VYNS build along with tools from OpenAI, Gemini, and xAI for various reasons I'll likely get into on my blog one day. The collaboration is genuine and unique. I am able to have the systems hold architecture decisions, debate tradeoffs, and generate production code. I feed different channels and chats live updates and make real-time decisions based on the current AI landscape from multiple angles including policy and regulation, new versioning, and tech releases. But as sessions grow, context fills and conversations end. Especially in Claude, to be honest, but that's because I'm most reliant on it currently. Each new chat starts cold and the discontinuity compounds over a long build. You rebuild context constantly, and the AI's knowledge of your project resets or pulls outdated info from relevant chats rather than accumulating properly and syncing only with updated decision frameworks rather than obsolete or pivoted ideas. Memory and past chat search help at the edges, but the fundamental architecture is still per-session.
LLM APIs can't read their own chat history, and that's a big gap
I structure my chats in Claude as departments for my startup, VYNS. I have a few chats that persist for personal use, and most other one-off chats I end up deleting after I gather the intel I need. But for VYNS, I keep a few persistent chats that are critical for my work as a solo founder. The most important of these are my Build logs, sequential chats where the previous one summarizes all key findings and stages as a prompt for the next to open and resume our work once the current chat reaches capacity. I also keep dedicated chats for Marketing, Philosophical, Competitor Research, Partnership Opps, and Local Resources. What I eventually want to build is an agent for each chat and an Executive Assistant agent that reads across and internally interacts with all of these based on my activity throughout the day, synthesizes findings, and reports back with a daily briefing highlighting my top priorities across the VYNS build and my personal brand. The limitation: no major LLM that I know of exposes read access to existing chat history. The threads are locked inside the web interface. This means I have to manually copy and paste references between chats and to my EA agent, as I largely do now. This isn't a small gap. It's a missing architectural primitive that would unlock an entirely new category of AI-assisted organizational design.