Scaffold Skills per AI Agent

bellatrix init places the same provider-neutral skills into each agent's native layout, and never rewrites a file you own.

Scaffold

Run bellatrix init in the root of your test repository:

bellatrix init --agent claude --domains desktop,web,mobile --json
bellatrix init --agent claude,cursor --dry-run --json      # several agents; preview only
bellatrix init --json                                      # re-run for the agents already scaffolded

Without --agent, init re-uses the agents already scaffolded (their manifests under .bellatrix/), else the single agent it detects from the repository, else it exits with code 2 and lists the choices. Re-running is safe: files you edited are skipped and reported unless you pass --regenerate.

Options

Option	Meaning	Notes
`--agent <a[,b,...]>`	AI agent(s) to scaffold for, comma-separated (windsurf = devin, agy = antigravity).Default: the agents already scaffolded, else the one detected from repo markers.	one of: claude \| copilot \| cursor \| devin \| codex \| gemini \| antigravity \| kiro \| opencode
`--target <a[,b,...]>`	Legacy spelling of --agent.	legacy spelling of --agent
`--language <language>`	BELLATRIX flavor. Default: the recorded language, else detected from build files, else csharp.	one of: csharp \| java \| typescript
`--domains <d[,d,...]>`	Domains to scaffold.	default: desktop,web,mobile; one of: desktop \| web \| mobile
`--out <dir>`	Repository root to scaffold into.	default: current directory
`--regenerate`	Overwrite files you edited (otherwise they are skipped and reported).
`--force`	Legacy spelling of --regenerate.	legacy spelling of --regenerate
`--dry-run`	Report what would be written; change nothing.
`--template-source <dir>`	Render your own template folder (same layout as the built-in templates) instead of the BELLATRIX templates.
`--prefix <prefix>`	Namespace for --template-source output (skills, agents, fragment, manifests). Default: inferred from the folder.

Also accepted: bellatrix update-skills (same as init).

Claude Code

bellatrix init --agent claude

Instructions	`.bellatrix/bellatrix.instructions.md` + one `@import` line in `CLAUDE.md`
Skills	`.claude/skills/`
Subagents	`.claude/agents/*.md`
Delegation	Isolated subagent (`@agent-<name>`)

GitHub Copilot

bellatrix init --agent copilot

Instructions	`.github/instructions/bellatrix.instructions.md` (`applyTo: "**"`)
Skills	`.github/skills/` (or reuses `.claude/skills/`)
Subagents	`.github/agents/*.agent.md`, with `agents:` allow-list on generators. Reuses `.claude/agents/` when Claude is scaffolded
Delegation	Subagent (VS Code `runSubagent`, CLI). Cloud agent: see limits

Cursor

bellatrix init --agent cursor

Instructions	`.cursor/rules/bellatrix.mdc` (`alwaysApply: true`)
Skills	`.cursor/skills/` (or reuses `.claude/skills/`)
Subagents	`.cursor/agents/*.md` (`readonly: true` on specialists)
Delegation	Isolated subagent (`/name`)

Devin Desktop (formerly Windsurf)

bellatrix init --agent devin

Instructions	`.devin/rules/bellatrix.md` (`trigger: always_on`)
Skills	`.devin/skills/` (or reuses `.claude/skills/`)
Subagents	`.devin/agents/*.md` (`allowed-tools`)
Delegation	Subagent (experimental). Fallback: `bellatrix <domain> find`

Codex CLI

bellatrix init --agent codex

Instructions	One marker block in `AGENTS.md` pointing at `.bellatrix/bellatrix.instructions.md`
Skills	`.agents/skills/`
Subagents	`.codex/agents/*.toml`
Delegation	Name the subagent in the prompt. Fallback: `find`

Gemini CLI (enterprise — Gemini Code Assist)

bellatrix init --agent gemini

Instructions	One marker block in `GEMINI.md` importing `@./.bellatrix/bellatrix.instructions.md`
Skills	`.agents/skills/`
Subagents	`.gemini/agents/*.md` (`kind: local`, Gemini tool names)
Delegation	Isolated subagent (`@name`)

Antigravity CLI (agy)

bellatrix init --agent antigravity

Instructions	`.agents/rules/bellatrix.md` (`trigger: always_on`)
Skills	`.agents/skills/`
Subagents	`.agents/agents/<name>/agent.md` (documented tool names; `mainAgent`/`subagent`)
Delegation	Subagent chosen by the main agent (no `@name` syntax). Fallback: `bellatrix <domain> find`

Kiro CLI

bellatrix init --agent kiro

Instructions	`.kiro/steering/bellatrix.md` (`inclusion: always`)
Skills	`.kiro/skills/`
Subagents	`.kiro/agents/*.md` (v3 Markdown; `resources` list the steering file and skills)
Delegation	Subagent named in plain language. Fallback: `find`

OpenCode

bellatrix init --agent opencode

Instructions	One marker block in `AGENTS.md` pointing at `.bellatrix/bellatrix.instructions.md` (shared with Codex)
Skills	`.opencode/skills/` (or reuses `.claude/skills/` / `.agents/skills/`)
Subagents	`.opencode/agents/*.md` (`mode: subagent`, `edit: deny` on specialists)
Delegation	Subagent (`@name`)

Rules every agent gets

Human-owned files are never rewritten. CLAUDE.md, AGENTS.md and GEMINI.md get exactly one import line or one marked block.
Skills are written once per directory set. When several agents read the same skills folder, BELLATRIX writes it once; --dry-run shows these decisions.
One manifest per agent. Re-running for one agent never touches another agent's files.
Delegation fallback is mandatory. Every fragment tells the agent to delegate locator work to the locator specialist or, when delegation is unavailable, to use only bellatrix <domain> find --snapshot, and never to read a snapshot or page source into its main context.
No MCP registration. init only removes BELLATRIX's own stale MCP entries; other servers' entries are kept.

Your own templates

Teams with their own conventions can render their own template folder (same layout as the built-in templates) with the same contract:

bellatrix init --agent claude --template-source ./my-templates --prefix acme --json

The prefix namespaces everything that is written: the skills and agents (acme-…), the fragment .bellatrix/acme.instructions.md and the manifests. bellatrix is reserved.