Skip to content

Agent skill

Lobbe’s skill is a SKILL.md file that teaches an agent how to work with Lobbe: send a meeting, wait for it, read its minutes and ask about it, through the CLI or the MCP server. It carries Lobbe’s rules too:

  • never print, log or repeat the token;
  • ask for a status no more than once a minute, and read the events feed for many meetings;
  • quote an Ask before asking it, and cap it at what the person agreed to;
  • send only what the person asked to send, since it spends their minutes;
  • at NAMING, give the person the naming link, and never guess names;
  • on DEVELOPER_MODE_OFF, tell the person instead of retrying.

Agents read a skill’s description to decide when it applies, so it loads when you ask about your meetings.

Terminal window
lobbe skill install

That installs it for Claude Code, for you. --target picks the agent, and --project puts it in the current folder’s project instead of your home folder:

Target For you With --project
claude-code (the default) ~/.claude/skills/lobbe/SKILL.md .claude/skills/lobbe/SKILL.md
codex ~/.agents/skills/lobbe/SKILL.md .agents/skills/lobbe/SKILL.md
cursor ~/.cursor/skills/lobbe/SKILL.md .cursor/skills/lobbe/SKILL.md
a folder <folder>/lobbe/SKILL.md

Any --target other than the three names is a folder, relative to where you run the command. The names match in any case; write ./cursor for a folder named like one of them.

  • Already there: an identical file is left alone (Already installed).
  • Something else there: a different SKILL.md is refused unless you pass --force. A symbolic link in its place is never written through: it is refused too, and --force replaces the link itself, never what it points to.
  • No token needed: the skill holds no secret, and installing it sends nothing to Lobbe.

lobbe skill install writes the skill that comes with the CLI you have. After updating the CLI, run lobbe skill install --force to replace an older copy.

The same text is served at /skill/SKILL.md, for an agent that reads skills from elsewhere or to save by hand:

Terminal window
mkdir -p ~/.claude/skills/lobbe
curl -fsS https://docs.lobbe.ai/skill/SKILL.md -o ~/.claude/skills/lobbe/SKILL.md
---
name: lobbe
description: Use when the user wants to send a meeting recording to Lobbe, follow it until it is ready, read its minutes (a "Lobbe"), search past meetings, ask about them, or look after their meeting series, vocabulary or memory. Works through the `lobbe` CLI or the Lobbe MCP server.
---
# Lobbe
Lobbe turns a meeting recording into minutes: a summary, decisions and action items, kept in the
person's archive. You act for one person through their personal API token.
## The loop
1. **Send** a recording: `lobbe upload <file> --json` (or the MCP tool `submit_meeting`). It
answers with the meeting's `id`, its status and the minutes it will spend.
2. **Wait**: ask for the status no more than once a minute (`lobbe status <id> --json`, or
`get_meeting_status`). The statuses go QUEUED → TRANSCRIBING → [NAMING] → GENERATING → READY.
3. **At NAMING**, Lobbe is waiting for the person to name the voices. Give the person the
`naming.url` from the status and let them do it in the app. **Never guess names.** If the person
tells you to go ahead without names, skip (`lobbe skip-naming <id>`, `skip_speaker_naming`).
4. **Read** it at READY: `lobbe get <id> --md` (or `get_meeting`).
5. **At FAILED**, retry once if `error.retryable` is true (`lobbe retry <id>`, `retry_meeting`);
otherwise tell the person what failed.
## Rules
- **Never print, log or repeat the token**, and never put it in a URL, a file or a commit.
- **Poll no faster than once a minute.** For many meetings, read the events feed instead
(`lobbe events`, `list_events`).
- **Quote before you ask.** An Ask spends minutes: get the price first
(`quote_ask_meeting`, `quote_ask_series`, or `lobbe ask` without `--yes`), tell the person, and
send the Ask with `max_minutes` (`--max-minutes`) no higher than what they agreed to.
- **Sending a meeting spends the person's minutes.** Send only what they asked you to send.
- A refusal with `DEVELOPER_MODE_OFF` means the person turned developer mode off in Lobbe's
settings: tell them; don't retry. `TOKEN_SCOPE` means the token lacks a permission: say which.
- Docs: https://docs.lobbe.ai (and https://docs.lobbe.ai/llms.txt).