recipe for successful PLANNING sessions in your software factory:
- use fable
- describe the rough shape of what you want to build in as much detail as possible, regardless if this is a fresh project or incremental addition (dictation works great)
- have fable pressure test/grill you about it: batches of questions to flush out the idea / edge cases / things you might not be considering
- dispatch research subagents (opus is fine) to ground or disprove your ideas with evidence, both from the internet and your existing codebase if there is one
- if this is a new project, draft a PRD. if this is an existing project that you are adding to, update your PRD.
- your PRD should contain whatever you think is necessary for your project. in mine, i generally like to have: project context, goals, non-goals, success criteria, key users/journeys, functional/non-functional requirements, key architecture decisions, data model & flow, contracts & boundaries, security & failure modes, open questions
- read and review the PRD (don’t skip this)
- set project-level cadence and policies. based on the shape of the project which you now know, what kind of testing makes sense? what kind of ongoing documentation makes sense? defining these now prevents random decisions by the agents in the future
- decompose your PRD into individual technical tasks (beads), optimizing for maximum parallelism during implementation, even if it’s optimistic and introduces a slight risk of merge conflicts, the speed makes it worth it imo
- technical tasks should be as detailed as possible - enough so that an agent with limited project context and only this task could still do it well. every task needs clear acceptance criteria
- after task decomposition and ordering, dispatch subagents (opus is fine) to perform adversarial review on your technical tasks, the mapping to the PRD, and anything else that could have possibly been lost in translation
planning like this is hard and takes a long time - and that’s the point, you are heavily front-loading the work. if you do this, your factory should be able to reliably execute very large batches of work either completely or mostly unattended (escalations/blockers only)
@momo5502 I use gvisor for my remote machine agent containers but nothing locally tbh not worth the trouble imo.
You have lots of cloud options though if you want disposable containers
0.9.0 is here, and it brings the most wanted herdr feature:
all your machines running herdr, in a single client 🎉
control your agents and projects across local and remote machines without jumping between terminal tabs.
working with AI is funny because you HAVE to develop your own systems to be a power user ... but then occasionally you need to just ask for the most ambitious /silly/almost naive thing you can think of that completely disregards all systems you've built just to remind yourself whats possible right out of the box