Chat commands
Type / in any chat composer to see the list. Commands either change how the next message runs (a sticky mode) or act once. Anything that isn't a command is sent to the agent as a task.
Reference
| Command | What it does |
|---|---|
| /plan | Plan mode: the agent proposes changes and shows a diff, but nothing is written to your repo until you accept. The default, safe posture. |
| /apply | Apply mode: edits land directly in your working tree as the agent makes them. Use once you trust the run. |
| /conduct | Toggle Conductor mode, or /conduct <task> for a single orchestrated run: the goal is planned, sub-tasks fan out to the model best at each, and the result is cross-reviewed. See the Conductor. |
| /one-shot | Toggle autonomous build, or /one-shot <brief> for one run: from a short brief the agent plans the full scope and builds it end to end, auto-applying. Runs longer than a normal turn. |
| /skill | Attach a learned skill by name (/skill lists them). A recorded skill replays its steps on send; a procedural skill guides the next message. |
| /brief | Scaffold a structured task brief — goal, acceptance criteria, constraints, test command — into the composer to fill in and send. A structured brief steers the agent far better than a one-line ask. |
| /clear | Clear the current conversation and start fresh in the same workspace. |
| /help | List the available commands inline. |
Plan vs. apply vs. one-shot
Plan and apply are about who accepts the edits. One-shot is about scope: it hands the agent a goal and lets it work autonomously until done. They combine with your chosen model and with Conductor mode.
- Plan — review every diff before it's written. Best for unfamiliar code.
- Apply — edits stream straight into the repo. Best once you're moving fast.
- One-shot — a brief in, a finished change out; the agent plans and executes the whole thing.
How skill replay works
When you record a skill, Lectern captures the concrete steps you took (the commands run, files touched, the recipe). Attaching that skill with /skill <name> and sending replays those exact steps rather than asking the model to re-derive them — deterministic, and it spends no reasoning tokens re-planning a procedure you've already proven. Skills also auto-apply when a task clearly matches one; the Hub is where you create, edit, and share them.
# attach a recorded skill, then send your message to replay it /skill deploy-preview