SpaceX AI employee just dropped the tunes for GROK BOT to make it reach its maximum potential.

already tried that myself and it shows AMAZING results. here is the list:

- add a self-check pass before it returns anything. ask the agent to verify its own output against the original instruction before finalizing, “did this actually satisfy every requirement, list what’s missing if not.” this single extra step catches more errors than any prompt wording tweak.

- define “DONE” explicitly. “manage my expenses” gets vague output. “flag every transaction over $200 with no matching receipt, format as a table” gets exact output. agents don’t guess well, they execute well.

- force structured output. ask for JSON or a fixed table format instead of free text. structured output is verifiable, free text just sounds confident either way.

- scope tool access per task, not globally. an agent with read access to everything “just in case” makes worse decisions than one with exactly the 2 tools this specific task needs. less access, sharper output.

which one of these are you actually doing right now?

SpaceX AI employee just dropped the tunes for GROK BOT to make it reach its maximum potential.

already tried that myself and it shows AMAZING results. here is the list:

- add a self-check pass before it returns anything. ask the agent to verify its own output against the original instruction before finalizing, “did this actually satisfy every requirement, list what’s missing if not.” this single extra step catches more errors than any prompt wording tweak.

- define “DONE” explicitly. “manage my expenses” gets vague output. “flag every transaction over $200 with no matching receipt, format as a table” gets exact output. agents don’t guess well, they execute well.

- force structured output. ask for JSON or a fixed table format instead of free text. structured output is verifiable, free text just sounds confident either way.

- scope tool access per task, not globally. an agent with read access to everything “just in case” makes worse decisions than one with exactly the 2 tools this specific task needs. less access, sharper output.

which one of these are you actually doing right now?

Source:https://twitter.com/elonmusk/status/2090924607889105014