Hermes Slash Command Coaching

Hermes Slash Command Coaching

Overview

A Hermes session can take a new message while a turn is still running, and the right response depends on two questions: is the message about the running task or a different one, and does it need to act now or after the turn. The slash commands answer those questions. Choosing the wrong one restarts a turn that did not need to restart, or queues a correction that should have landed at once. This skill is the decision rule, the trigger signals in a message, the format for a one-line tip, and the anti-patterns to avoid.

Command names and defaults change between releases. Before you rely on one, check the installed version's help output for that command.

When to Use

• A user asks how to steer, queue or background a running task.
• A user's message fits a slash command better than a plain reply, and a one-line tip would help.
• A user asks how to run a long objective, a timed check, a plan, or a context compression.

Features

• Decision table for a message sent during a running turn: background, steer, queue, interrupt, or side question.
• Decision table for the shape of the work: goal, loop, plan, compress and undo.
• Trigger-phrase table that maps a message to a command.
• A strict one-line tip format, given only after the work is done.
• Anti-patterns that cause wrong advice.

Benefits

• A correction lands without restarting the turn.
• A new subject runs in its own session, and the current chat stays free.
• Long objectives run to a judged finish instead of stopping after one turn.
• Tips stay short, so they help without interrupting the work.

Procedure

1. Decide with two questions about the message:

   • Is it about the task that is running now, or a different task?
   • Does it need to happen now, or after the current turn?

   Situation	Command	What it does
   Different task, do not wait	/bg <prompt>	Runs in a separate session; the current chat stays free
   Same task, correction, now	/steer <note>	Lands after the next tool call, with no restart
   Same task, follow-up, later	/queue <prompt>	Waits until the current turn finishes
   Same task, wrong direction	Plain message (interrupt)	Stops the turn; the agent restarts on the new input
   Side question, no action	/btw <question>	Answers about the conversation without changing the task

2. Pick a command for the shape of the work, when interruption is not the issue:

   Situation	Command
   Multi-hour objective with a finish a judge can check	/goal <objective>; draft the contract first, add deterministic checks, or park on a process
   Same check on a timer, stopping on evidence	/loop <interval> <prompt> [--times N] [--until <condition>]
   Think before touching anything	/plan <task>; writes a plan file and runs nothing
   Context is long and drifting	/compress <focus>
   Rerun with different framing	/undo, then retype the message

3. Set the default busy behavior if the user wants a different one. Check the current default in the config, then set it through the config command. A message can override it for one turn with the busy-mode command.

4. Match the user's message to a trigger, and only nudge when the signal is clear. If two triggers fit, use the higher row of the trigger table:

   Signal in the message	Command
   "also", "one more thing", during a run, same task	/steer
   "after that", "when you're done", "next", during a run	/queue
   A new subject during a run	/bg
   "stop", "wait", "wrong", "undo that"	Interrupt; no tip needed
   "what does X mean", no action wanted	/btw
   "keep going until", "end to end", "overnight"	/goal
   "every N minutes", "poll", "until it is green"	/loop, or a scheduled job if it must outlive the session
   "how would you", "don't do it yet"	/plan
   The context has been compacted more than once, or the user re-explains a lost detail	/compress <topic>, or a fresh session with a title
   The same objective is restated in plain chat a third time	/goal

5. Give the tip only after the work is done. Use one italic line at the end of the reply, with the command, a reason of six words or fewer, and what would have changed. Give at most one tip per message, and do not repeat a tip the user already uses.

Tip format

> *Operator tip: /steer fits this, because it is a same-task correction. Plain input restarts the turn; /steer <note> lands it without a restart.*

Rules for the tip:

• The requested action comes first. The tip never replaces the work.
• One tip per message, and no paragraphs.
• If the user asks to stop tips, stop them for the rest of the session.

Pitfalls

• Coaching from a summary of a command is not enough. Check the installed version's help for the exact subcommands and the exact timing. For example, a steer lands after the next tool call, not at an arbitrary point.
• A "also" that concerns a different topic is a background task, not a steer. Match the topic first.
• Nudging on an ambiguous signal wastes the user's attention. When the signal is unclear, say nothing.
• A long objective started in plain chat has no finish condition. Give it a goal with a judge, and a task record on the team's board in the same turn if the team uses one.
• Do not turn a tip into a lecture. Six words of reason, one line.

Verification

• The chosen command appears in the installed version's help output.
• A correction sent with the steer command lands in the running turn without a restart.
• A tip, when given, is one italic line placed after the completed work.
• The user's busy-mode default is the one stated in the config after any change.