kickoff

Imported from youngxguo/dotfiles (claude/skills/kickoff/SKILL.md). Install upstream with npx skills add youngxguo/dotfiles --skill kickoff. Copyright stays with the author.

Kickoff

kickoff.py beside this file runs the whole sequence: resolve the repo's main checkout workspace, herdr worktree create a branch and open it as its own workspace, wait for the root pane's shell, ask the rebump skill which account and model have headroom, launch claude there, answer the folder-trust dialog a fresh worktree path shows, name the agent, and send the brief.

Needs herdr on PATH and the rebump skill beside this one; this session should have HERDR_ENV=1.

Use it

python3 ${CLAUDE_SKILL_DIR}/kickoff.py signup-rate-limit --brief-file /tmp/brief.md

The slug is the only required argument. It becomes the branch (<git user>/<slug>), the workspace label and the agent name. Write the brief to a file in the scratchpad and pass --brief-file; only use the trailing positional form for a one-liner.

Worth knowing:

• --repo <name> kicks off in another repo's main checkout, so this works from any pane. Without it the source is the current workspace, resolved up to the repo's parent checkout if you are already inside a linked worktree.
• --base <ref> branches off that ref instead of HEAD. In a top-level workspace HEAD is usually what you want.
• A branch that already has a worktree is opened rather than created, and the run stops if an agent is already living in it (--force overrides).
• --focus switches the user to the new workspace; the default leaves their focus alone.
• --no-agent just opens the worktree workspace. --kind codex starts a different agent (rebump only applies to claude). --to c3 pins the account. --wait blocks until the agent settles on the brief instead of returning as soon as it is sent. --trust-repository is for a repo herdr has not been pointed at before. Anything after -- is passed to the agent itself.
• --dry-run prints the branch and source workspace it resolved and stops.

Report the final line it prints: agent name, branch, workspace id, path.

Write the brief like the agent has no context

It starts fresh in an empty worktree and cannot see this conversation. A brief that works is a short paragraph or two covering: the goal in one sentence, the files or entry points you already know are involved, the constraints that are not obvious from the code, and what done looks like - tests to pass, and whether to commit, push and open a PR. Say that its branch is already created and checked out. Do not ask it to write its answer to a file; read the pane.

Afterwards

Check on it, or send it more work, with the herdr skill's agent list, agent get, agent read and agent prompt - the agent answers to the name this printed.

A finished worktree is cleaned up with herdr worktree remove --workspace <id>, which closes the workspace - only on an explicit ask, and never for a workspace this skill did not create.