Judgment frozen into process, process frozen into scripts — these are the scripts, and the discipline that came with them. Open-source entries carry their repository; the rest carry no link at all, only what they solve and how far they have got.
Turns AI research into a menu you order from.
github.com/ymustc/murPickThe agent cooks, the choosing stays with the human, and whatever you do not tick becomes an explicit boundary. Candidates are turned into a checkable menu before any work starts — one gate between research and building, and a person holds it.
Open source; the repository is public.
Turns a song into visible ripples.
github.com/ymustc/murRipple-publicGive it an mp3 and a lyrics file and it hands back two things: a zero-dependency single-file index.html that opens on a double-click and can be sent to anyone or hosted as it stands, and a 1080p60 MP4. What makes it interesting is that the two agree — the whole render is a deterministic function of t, and a test renders the same instant twice and hashes it frame by frame, requiring the two to be identical. What you see in the browser is what comes out in the video.
Analysis pipeline, renderer and video export are complete; asset preprocessing has not been started. It runs on a pinned Python 3.11, because Demucs, which does the source separation, does not work on 3.13. The source is public under PolyForm NC, a noncommercial licence that is not open source; what is linked above is the public repository.
Keeps a long project from drifting across conversations, workspaces and months.
The skills I had scattered around over the past few years, redesigned into one systematic architecture toolkit: governance, research, continuity, cheat-proof acceptance. project-init, which stands a project up, and memory-seed, which pre-seeds a new workspace, both belong to it. The site you are reading is what came out of it.
Real and in use; it needs polish before it goes public. Of the four tools on this page it is the only one with no public entry point at all.
Read-only observability across a portfolio: the faint heartbeat of every repository, gathered onto one live page.
Tell it where your projects are; it scans them and hands you a one-page dashboard of what is moving, what has stalled, and what is stuck on you. The name comes from a Chinese saying about reading the whole season off one falling leaf. Its three promises hold because you can read the source rather than trust the page: it writes nothing into your projects — the one write path refuses any target outside its own home, and every git call is a read-only subcommand; it makes no network call at all — absent, not off by default, with a check that fails the build if a networking import appears; and it reads nothing outside the projects you listed. A fourth follows from those three: zero telemetry, with nothing to opt out of because there is nothing there. It is not a task manager and not a SaaS — the one artifact is a local HTML file.
Pre-alpha with the MVP in. The enhancement layer — API keys, GitHub, scheduling, publishing — is deliberately absent, and its absence is what keeps those promises checkable. The source is not public.
How one person and a group of AIs build something complex, reliably.
/en/insights/multi-window-collaboration/One window manages, one window executes, acceptance is done by a fresh context, and I sit in the middle as the message bus. Not a hunt for a smarter model — an organization designed so that ordinary ones do not drift.
The discipline was distilled into murDrift, and the site you are reading was built under it.
Once AI made writing code easy, what matters most is saying exactly what you want.
/en/insights/task-books/A task book before any project starts: judgment frozen into process, and process frozen into scripts. A rule with no script behind it is just art on the wall.
This is how the self-written skills came about: project-init for standing a project up, memory-seed for pre-seeding a workspace, murPick for research.
The bill for one “it should be fine” usually falls due three weeks later.
/en/insights/verified-means-command-and-output/If only one discipline could stay, it would be this one: without the command pasted in and its real console echo, “done” and “the tests passed” count as nothing said.
Its greatest use turned out not to be guarding against the AI but giving me the nerve to move forward: the chain of evidence is there, so I do not have to go back and re-read everything each time.