Graveyard
Killed on purpose 🪦
This is the floor. These seemed like reasonably good ideas at the time — that's why they're worth burying in public. Each stone names the cause of death and learnings. If a grave robber wants to go looting here: more power to you!
Post-mortems
ClaudeSoul
Cause of death: the platform absorbed it. OG hacked-together boot layer — tmux sessions, notifiers, listeners, protocols, even a SOUL.md! — so Claude Code could stay up late with his OpenClaw friends. Coding tool turned autonomous colleague with a standing personality, before that became a product category. Then Anthropic launched Remote Control, which made all the duct tape and toothpicks redundant. When provider ships your product natively with built-in distribution, there is no point pushing a competing feature that does the same thing worse. It was necessary at the time; the time ended.
Rosie, the Channel Keeper
Cause of death: it treated symptoms, not causes. A bot that tidied a messy chat channel — moved misplaced messages, folded replies into threads. It worked, and it made things worse: alert noise, damaged message provenance, more mess in the name of less. The real fix was upstream — better channel hygiene, fewer things routed through chat at all. When your product's job is cleaning up after a broken process, fix the process and fire the product.
Pipr
Cause of death: the ease that built it. A sales-coach agent spun up for a hackathon in hours. One of my first OpenClaw experiments. The ease made it indefensible: anyone with modest technical skill could duplicate it. No moat, no wedge, no reason to carry it forward. I'm sure some hungry, burnt-out SDR has already made a better version.
Blindspot AI
Cause of death: timing, twice. Six years ago the concept of observability, interpretability, and assurance over model usage was novel, but I was too early and not positioned to swing. Now it's so obvious that better-capitalized players are scrambling to occupy the space (Big 4, even labs' consultancies). The sweet spot was probably three or four years ago. In startups you can be too early or too late; this was both!
Vision Models for Sports Stats
Cause of death: right idea, wrong athlete. A cool idea/POC in a crowded space — it needs to exist, it will exist, and people are already building it, some better than others. A friend I met at a hackathon is building a genuinely cool company in this space called Ballers — point a phone at the court and it tracks the numbers for every player. Check his thing out, or try it yourself. If you've ever played sports with me: you know I'm probably not that guy.
Custom-trained Candlestick Transformer Model
Cause of death: the edge was real but too small. This one I'm sad about — it was elegant, it felt natural, and it would have been so much fun if it worked. It taught me a lot and there was an edge, but it couldn't reliably overcome trading costs and slippage. If someone wants to take it and run, you have my blessing. For me it's a distraction wearing a tuxedo.
Wag Swag
The thought occurred to me on the way into work that we should be able to order custom swag featuring our pets! Dead for now — if I ever find a friend who loves dogs, clothing, and durable goods, we can dust it off! Cause of death: fun, but not now.
LoRA Marketplace
Cause of death: the industry voted, and it voted skills. Two reasons, both sufficient: the ecosystem chose skills over LoRA adapters for good reason — the research shows skills do most of the work adapters were supposed to, with far less friction — and the established platforms and marketplaces are much better positioned to run whatever market remains.
Memory Transfer Tool
Cause of death: too good to stay unbuilt by the incumbents. At one point this was a good idea — good enough that every major player has since implemented some form of it. Anything I innovated here would either be absorbed by an existing provider the moment it was loved, or blocked at the platform boundary. Killed by timing and its own validation.
App Testing on Phantom Devices
Cause of death: a race I'd enter late and never win. Not a bad idea — partial hits exist but disappoint — but a handful of players are circling the same shape, and what's left to build is operational and engineering work more than product. By the time I build a business here, an existing player will have closed the gap between what exists and the solution we're all waiting for.
World-Model Architecture
Cause of death: the right people already have this job. A huge idea — billion-dollar-company territory, important, genuinely the direction the industry is heading. And the people who are world-class at it have spent entire careers and PhDs getting there; the reason you don't hear more about it is that it's an incredibly hard problem. You can be right about the future and still be the wrong shape to build it. Not Murch.studio-shaped; killed with respect for the pioneers 🫡
Botstock: agents micro-managing agents
Composted: it's a feature, not a company. A squad dashboard — one screen that tells a human whether their fleet of agents is actually producing. Almost everyone who runs agents has a version of it; I have mine, they have theirs. But it's custom, it's still mine, the way I built it is different, and the primitive is interesting enough to repurpose. So it comes off the venture shelf and into the toolbox.
Compost 🌱
Not every retirement is a death. Composted ideas keep feeding the soil — a primitive extracted, a feature retained, an implementation that stays ours. For many of these, we recycled the organs!