npm install -g oyeboss # the usual — macOS, Linux, Windows brew install ajeshh/boss/oyeboss # macOS; then plain `brew install oyeboss` works # then open the project in Claude Code — terminal, VS Code, Cursor or JetBrains, your call
/plugin marketplace add ajeshh/bossbuild # the repo is its own marketplace /plugin install boss@bossbuild /boss:welcome # the front door — asks what you're building, says it back, then offers to set it up
The plugin is deliberately thin: one skill and boss on Claude's PATH. Everything
else (the mentors, the conscience, the skills) arrives in your project when you say go. It
is not yet listed in Claude's plugin directory; the lines above add this repo directly, and
that path is verified against the current release.
Something new
Scaffold it
boss new my-app creates the project, git-inits it, registers it, and starts
you in Quickstart with 4 agents and 15 skills.
Open Claude Code
cd my-app then claude, terminal or editor panel. First time?
/welcome is about a minute, and it asks where you're starting from so the tone
fits.
Point it at the idea
/boss <your idea>: a sentence, a doc, a Google Doc, an Obsidian note, a
PDF, or a URL. It says the idea back in your words first (what it is, who it's for, the
smallest version that proves it, and which way it read any word that could go two ways),
then asks three things nobody else does: why this one, for you, what would
“it worked” look like in three months, and if it works, what's here in a few
years. Skip is an answer. Already written it down?
/import <file|url>. Just want to see something move?
/prototype builds the smallest clickable version to react to.
Come back later and run boss status
The one command worth remembering. After a few days away it tells you where you are and what's next.
A repo you already started
Don't start over. cd into it and run boss adopt; it reads how far
along you already are (a build manifest, source files, tests, CI, a deploy config), prints
what it found, and starts you at the mode that matches.
Non-destructive throughout: your files are untouched, and an existing CLAUDE.md
or AGENTS.md gets a marked block appended rather than replaced. It
caps its own guess at MVP on purpose, because V1 is a design-system and database
commitment, and that's your call, not a machine's. If it reads low,
boss unlock <mode> is one command.
Then run /read-repo inside Claude. That's the adopted repo's version of
/boss: BOSS reads the actual code, tailors the scaffold, and seeds the venture
brain so the conscience starts with real context instead of guessing.
It tunes itself to you
/welcome asks this once, in these words (four of its eight answers). It changes how much it explains and how the conscience frames a nudge, nothing else; the mode is yours.
/idea; the payoff is /canvas. Unlock MVP when you're actually
building, not before.
/ai-failure-states
as soon as your code first calls a model, and /red-team --humane before anyone relies on
it.
Changed your mind?
boss remove previews exactly what would go; --apply does it. It only
removes what BOSS wrote; your code, the ideas and decisions you captured, your own
skills, and any BOSS file you edited all survive, and your CLAUDE.md keeps
everything except BOSS's marked block. boss remove --global takes it off the
machine instead; your projects keep working, since the conscience hook runs from the project,
not the CLI.
And it hands back the thinking on the way out. If BOSS built up a venture brain about your
project (what it understood, what it concluded) that gets exported to
docs/venture-brain.md before anything is deleted. You wrote into it and
you could edit it, so it leaves with you rather than with the tool.
A clean exit is what makes the entrance safe to try.