One sentence before bed, results by 8: She split one Bot into 15

SpaceXAI engineer Lauren Tan has publicly shared her complete workflow for running agents. The core is one sentence, sent to Grok Bot before bed:

“I’m going to sleep. Land the stack, even if CI is unstable.”

After this sentence, Grok Bot does not work single-threaded. It is organized into a graph of 15 bots; the task is sliced, advanced in parallel, and cross-verified, and by morning it pushes everything it can into a reviewable state. She posted a demo video:

In the post, she lists seven actions for this prompt: first tighten the requirements and write pass conditions; slice the task and run it in parallel; have other models try to break every diff, keeping only what the run can prove; check what a one-line change might still break; remove comments and AI traces before sending for review; turn each repeated error into a rule for the next round; and wake the human only at irreversible steps.

This workflow solves an old problem with agents: the vaguer the task, the more the bot goes looking for work on its own, spending your tokens to explore until it feels it has done enough. Lauren Tan’s approach is to define “done” clearly, then split the big task into independently verifiable slices. The result of each slice is not proven by the bot saying “finished” but by the run output.

Another background made this noteworthy. Developer maestro wrote a long post in the same thread: he burned 99% of Grok Bot’s weekly quota in three days, and 84% of the tokens went to the bot rereading its own chat history. Among his 12 methods, the most direct is “start a new conversation for each task.” Because every round the bot rereads the whole chat, that conversation that has been piling up chores since Monday gets more and more expensive.

He also mentioned the frequency of scheduled tasks. Triggering every 15 minutes is 96 times a day; triggering every hour is 24 times; three times a day is only 3. For most tasks like “check the inbox,” three times a day is enough. The same goes for event triggers: if a Slack trigger is set to “any message in all channels,” even memes will wake the bot. It should point to one channel and one phrase.

Spending caps are worth setting. But maestro warns that a cap will not cut off a task mid-run, and a task already started may go slightly over. A more reliable approach is to put long tasks into Custom Rules so the bot asks you before it starts.

Task briefs also need to be tightened. He gave a template:

Task: [one sentence]
Done means: [the exact output I will check]
Use only: [allowed apps, folders, or websites]
Stop and ask me if:

• It needs something outside the list
• The same step fails twice
• It has gone more than [N] steps without getting close to done

This is the same idea as Lauren Tan’s “write pass conditions first”: put “what counts as done” up front, and the bot will not use your money to guess.

maestro’s full post has a few more points: give scheduled tasks a separate bot, so all scheduled runs do not pile into the long conversation you use for daily chat; use the smallest real input when testing routines, because tests do touch real data; and turn off app tools not needed for the current task, so the bot has fewer entrances for wandering.

If you want to try it, start with a small task. Before bed, write the task, completion criteria, and allowed scope clearly, and send it to a separate bot. The next day, look at the result first, then at how many steps it took. Saving money and efficiency are the same thing here.