A shape that won't settle
Of all the small tools I've been writing, this one is the one I've rebuilt the most. Five or six times, depending on what I count. Each rebuild fixes a real friction. None of them have produced the version I keep using. I'm writing about it now not because I've found the form, but because I haven't, and the not-finding feels like part of the work.
The itch is recognisable enough to keep returning to. I want to coordinate Claude agents across more than one step, in a way that doesn't bury me in their state. Numbered prompt files, dependencies between them, parallel runs that don't step on each other, named roles I can re-use. The shape on paper is clear. The shape in my actual week keeps slipping.
Drafts and dodges
The first form was a directory of numbered prompt files with a depends_on field, walked as a graph in dependency order. It still works exactly as intended. What I noticed about myself, after building it, was that I'd start a chain, run the first prompt, end up in a follow-on conversation about its output, and never run the second prompt through the tool at all. The graph was correct. My way of working with it wasn't.
The second form added named personas with scoped tools and an explicit system prompt. The idea was that the parts of my work that recur could become re-usable subagents. What I noticed was that the personas I most wanted to keep were the ones I wanted alongside me while I worked, not spawned as a subprocess. The tool could emit them as host-session prompts (one of two invocations), but at that point the value was mostly in the file, not in the runtime. The clearest sign of that came later, when I added a mode: field to the persona contract: a role-shaped persona invoked as a one-shot subprocess would silently produce a transcript asking for a reply nobody was sitting there to give. The fix was a few lines of YAML and a guard rail. The runtime around the file kept shifting; the file shape stayed stable enough to be worth annotating.
The third form added shallow-clone profiles so parallel runs could be isolated without me having to spin up fresh Claude config dirs by hand. Solved a real problem. What I noticed was that most of my parallel work happens through tmux panes with hand-started Claude sessions anyway: one main, a couple of throwaways. The garbage-collected ephemeral profiles are over-engineered for what I actually do most days.
The fourth form added terse mode, a compression instruction that holds the agent's output tight on long runs. Useful when you're running a long pipeline. I rarely run a long pipeline, because I'm rarely running pipelines through the tool to begin with. The optimisation is real. The use is what's missing.
There are smaller forms I've tried alongside these (a feature-and-project workspace layer, a global daemon for notifications, a lead-session entry point). Each adds a piece. Each fits a real shape of work. None of them are the shape that turns the tool into something I open by default.
The deeper miss
What I notice, looking across all of these, is that each form fixes the friction it set out to fix and doesn't move the deeper one. The deeper friction is that the work I most want a tool like this for is also the work I most want to be in. I keep designing the tool to take me out of the loop, and I keep wanting to be in the loop. The other recent post here was about that exact thing in a different domain. The orchestration tool runs into it from the other side.
The work that genuinely fits the tool's shape (release notes, inbox triage, scheduled background passes) is real, but it isn't enough of my week to drive the design. I keep building features for the cases I wish I had more of, and ignoring the case I actually have more of, which is one prompt at a time inside a single session.
Again and again and again
I keep coming back to it for a couple of reasons. The underlying urge hasn't gone away, even after several rebuilds. There's a class of work where I want orchestration and don't want to be present, and that class is going to grow if I let it. The tool I'd want for that work doesn't exist yet, even in my own attempts.
I also notice that each rebuild has produced one piece worth keeping. The numbered-prompt format is good. The persona file is good. The shallow-clone integration is good. They don't add up to a tool I open daily, but each one is a primitive I'd pull off the shelf if I found the right work to point them at. I haven't given up on the integration; I've just stopped pretending I'm close.
What I'm sitting with now is the possibility that the right form is much smaller than what I've built, or much larger, or doesn't look like a CLI at all. Maybe a recipe pattern, maybe a daemon I never invoke directly, maybe a thing my host session calls into without me thinking about it. I don't know. I'm willing to rebuild it a sixth time and a seventh if that's what it takes, because I'd rather sit with the not-finding than declare a form I don't actually use.
The next post in this arc is about the one piece I do think landed cleanly: the persona file with two invocations. That distinction came after I'd been running both kinds for a while and noticed they were different in a way the tool could name. It's a smaller claim than "here's the orchestration layer," and it might be the only claim worth making about this tool right now.