Best practices

#Treat it like a junior teammate

Fast, tireless, willing — but needs briefing. Set context up front, give access to the right tools, then let it work. Don't spoon-feed each keystroke, and don't treat it like a finished product either.

#Verify before trusting

• "Show me the URL you used" / "show me the file you edited" — concrete artifacts beat the assistant's word.
• "Did you actually check, or are you guessing?" — direct, useful question.
• For anything that touches money / contracts / external sends — assume it might be wrong and confirm.

#Write things down

#Use sessions like notebooks

One topic per session is ideal. Long-running threads degrade context and eventually crash the message size limit. Better:

• Use the same session for related work in one sitting
• Start a fresh session when switching projects or starting a new day
• Before closing a session, ask the assistant to "write a handoff note to today's daily log"

#Build dashboards for state

For state info — what scheduled jobs are running, what projects exist, what permissions the assistant has — build (or use) a dashboard instead of asking the assistant. Faster, more accurate, saves context for actual conversation.

#Push back

Assistants default to agreeable. If you suspect it's wrong, say so — "I don't think that's right." It'll either defend its position with reasoning or admit it was guessing. Both outcomes useful.

#Use external tools where the LLM is weak

• Math → ask it to write code that computes it
• Counting → same, use code
• Dates / day-of-week → ask it to print today's date first
• Visual tasks → screenshot + describe what you see
• Up-to-the-minute info → tell it to search the web

#Brief at the start of complex work

Don't just dive into "build me X." Set the scene first:

• What you're trying to achieve (the goal, not the immediate ask)
• What you've already tried / decided
• What constraints are non-negotiable
• What success looks like

#Time-box

Without a time-box, the assistant will keep going. With one, it prioritizes. "Get me to a working v1 in 30 min" produces different work than "build the best version possible."

#Be explicit on external actions

When the assistant drafts an email or social post, you'll see "send it / edit / skip." Be explicit. Ambiguity here is where things get accidentally sent that shouldn't.

#Be okay with imperfect first drafts

Most outputs are better as a 2-iteration thing. First pass = scaffold + rough content. Second pass = your edits + tightening. Don't try to one-shot important things.

#Audit periodically

• Every week: "what scheduled jobs are running? Am I still using all of them?"
• Every month: "what's in MEMORY.md that's stale and should be cleaned up?"
• When something feels off: "show me your recent admin actions" / "what emails did you send in the last 24h?"

#Keep it away from things it doesn't need

Don't feed an assistant secrets it doesn't need. Don't share private docs "just in case." Smaller exposure = smaller blast radius when something goes wrong.

#When stuck, ask what it'd do