$ /venture "an idea"
Venture case study · in progress
A Claude Code skill that takes a raw idea to a store release, and says no when it should. One orchestrator, phase files loaded one at a time, subagents doing the work, a human at every gate that matters.
- version
- v0.3.1
- entry points
- bootstrap · resume · adopt · retro
- runs to store testing
- 2
- lessons folded back
- 58
$ cat SKILL.md | grep phase
The pipeline
- intakethe idea, parsed
- validationhonest, incl. why it fails
- gatego · pivot · park
- wizardname, stacks, gates, infra
- researchmarket & tech findings
- setuprepos, CI, deploy
- planningepics into BACKLOG.md
- epic loopTDD, design loop, PRs
- releasesigning, listing, testers
- retrolessons back into the skill
The epic loop is the part that runs unattended: pick the first unchecked epic, split it into sprint tasks, route UI work through a design loop, execute each task through TDD on its own branch, flip the checkbox when it lands. Kill the session and it resumes from the two state files.
$ cat SKILL.md | grep rule
Hard rules
- menus, not guessesevery technology choice is a menu with one recommended option, never a silent pick
- the gate always stopsvalidation ends in go / pivot / park and waits for a human, whatever the automation setting
- two files of statePLAN.md holds decisions, BACKLOG.md holds checkboxes, nothing else remembers anything
- publish needs consentrepo visibility, store submissions and DNS ask again at the moment of execution
- no leaked identitye-mails, git names and org names are never inferred from the machine
$ tail BACKLOG.md
What state looks like
$ ls ~/ventures --shipped
Built through it
$ ./write case-study --format article
▸ drafting
The full write-up (decisions, retro lessons, what the agents got wrong) is being written. This page grows into it.

