nvardakas

This is great. There's something refreshing about discovering websites by accident instead of through an algorithm optimizing for engagement. I've been spending more time on personal blogs lately and the writing quality is consistently better than anything on Medium or Substack probably because nobody's writing for SEO.

This is the part that doesn't get talked about enough. Code review was never just about catching bugs it was how teams built shared understanding of the codebase. When someone skips reviewing their own AI-generated PR, they're not just shipping unreviewed code, they're opting out of knowing what's in their own system. The trust problem isn't really about the AI output quality, it's about whether the person submitting it can answer questions about it six months from now.

I haven't tried smaller models for multi step pipelines specifically, but my intuition is that splitting into 4 steps with a 27b model is going to hit the same coordination overhead problem just with worse reasoning at each step. The wins I've seen with smaller models are on single well scoped tasks formatting, extraction, classification where you don't need the model to hold a complex plan in its head. For anything that requires understanding why you're doing step 3 based on what happened in step 2, I'd rather use one frontier call with good context than four smaller ones.

Right, better architecture, cleaner code is the AI equivalent of synergy in corporate emails. It sounds like it says something but it communicates nothing. The useful PR description is changed X because Y was breaking Z and that requires the author to actually think about what they did. If the tool is doing the thinking, the description is just decoration.

The expert systems parallel is a good one. The difference this time is that the composition engine actually understands intent rather than just following hardcoded rules, so the deterministic pieces can be much smaller and simpler. You don't need the expert system to handle edge cases the LLM routes around them and the deterministic tools just execute. Re: trajectory view of agents I haven't read the COALA paper but the framing makes sense. Agents are clearly path-dependent in practice the same prompt produces wildly different outcomes depending on what tools fired in what order earlier in the session.

Honestly I'm still figuring it out. The main shift was stopping all the broadcast stuff tweets, directory submissions, cold DMs and just going to where people already talk about the problem I'm solving. I ask open ended questions without linking anything. If what I'm building is actually useful, people describe the pain themselves and sometimes ask what I'm working on. The conversion is way lower volume but the conversations are real. The hard part is patience it feels like you're doing nothing compared to the dopamine of posting a launch.

The control center apps are always the worst offenders. It's bizarre that a company can nail the hardware engineering and then ship companion software that feels like it was written as an intern's first project. At least with open source alternatives like Mouser you can just bypass the whole thing. I wonder how many people have switched mice entirely just because of bad software.

The task.md as a programming-language idea is interesting. I've been doing something simpler just a CLAUDE.md with key decisions and project context but the gate pattern sounds like it would solve the problem I keep hitting: the model drifts from the plan halfway through a long session even when the context is all there. Having explicit checkpoints it has to pass through would force it to stay on track rather than just having the information available and ignoring it. How do you handle when a gate fails does it retry the previous step or do you intervene manually?

Same thing with PR descriptions. The signal-to-noise ratio has completely flipped. Before, a short PR description meant the dev was lazy. Now, a long detailed one might just mean they hit generate description and didn't even read it. The length went up, the usefulness went down, and the reader has no way to tell which kind they're looking at.

Very good catch. The mental model thing is real I've caught myself approving LLM generated code that works but that I couldn't debug if it broke at 2am. With libraries you at least had docs and a community. With generated code, the only source of truth is... asking the same LLM again and hoping it's consistent.

This matches what I've seen too. I spent time building multi step agent pipelines early on and ended up ripping most of it out. A single well prompted call with good context does 90% of the work. The coordination overhead between agents isn't just a cost problem it's a debugging nightmare when something goes wrong and you're tracing through 5 agent handoffs.

Same situation here. English is not my first language and I use Claude constantly to clean up my writing. Not to sound like someone else, just to make sure what I'm trying to say actually comes across clearly. The irony is that the more polished text sometimes gets less engagement because people assume it's fully AI generated, while a messier version with obvious non-native patterns would feel more human.

This resonates. The pattern I keep seeing is that the best AI tooling right now is about constraining the agent, not giving it more freedom. MCP gives you a clean boundary between what the AI decides and what the system executes deterministically. I use MCP servers with Claude Code and the biggest win is exactly what you described, the AI handles the creative problem solving but the actual actions go through predictable, auditable paths.

Can confirm this from direct experience. I spent 3 weeks doing everything on the checklist Product Hunt launch, daily X posts, shared in communities with 250k members, cold DMs, directories. Result was literally zero signups. Not low, zero. The noise level is so high now that the standard playbook just doesn't work for indie developers anymore. What finally started getting real feedback was asking genuine questions in relevant subreddits instead of promoting actual conversations instead of broadcasts.

The fact that Logitech ships hardware this good with software this bad is wild. It's like they have two completely separate companies. I switched to just using the onboard memory profiles on my mouse and never opening the app not ideal but at least nothing is eating 40% of my CPU in the background.

This is so true. The algorithm optimizes for engagement, not discovery. I find myself watching the same types of videos over and over. Tools like this that bring back the randomness of channel surfing are a nice antidote you stumble into things you'd never search for.

I've found the same thing. I build with Claude Code daily and the context decay is real by the end of a long session it starts forgetting decisions we made earlier. The 1M context window should help but I'm curious how coherence holds up at that scale.

What's been working for me is keeping a CLAUDE.md file in my project root with key decisions and context. The model reads it at the start of every session so I don't have to re-explain everything. Not as elegant as automated compaction but it works.