Plain input - a one-off nudge
Type a message with no slash command and it’s ephemeral: injected into the current iteration only. The next fresh iteration doesn’t see it. AutoPT does not sort your message into anything more durable on its own - a plain prompt is always a one-off. Use it for small, in-the-moment corrections:- “Slow down and confirm that finding before moving on.”
- “Ignore the staging subdomain for now.”
- “Try a different user-agent on that endpoint.”
Making guidance stick - the slash commands
Three commands give your input a longer life than a single iteration./pin <text> - a standing note
A pin is re-injected into every iteration’s prompt and shown to the agent as authoritative pinned context. This is where durable facts belong:
- credentials and tokens the agent should reuse
- leaked or hard-to-find URLs
- scope constraints (“never touch
prod-db”) - any key fact you don’t want re-discovered every iteration
/unpin <id> - remove a standing note
Pins are numbered in the order you add them (1, 2, 3…). Remove one by its ID:
/pin command writes to.
/do <task> - a one-shot task
/do hands the agent a single task, queued at high priority. It runs once, gets logged, and is not repeated. Use it to redirect attention right now without leaving a standing note behind:
Which channel do I want?
A good habit at the start of a run:
/pin the facts that define the engagement (credentials, scope limits, known URLs), then use plain input and /do to steer as you watch the thinking stream.
