Write a brief a Grok Bot can actually finish

A Bot is only as finished as its brief. Most “it went off and did something weird” stories trace back to a request with no finish line, no named sources and no point where it had to stop and ask.

The five parts xAI recommends

The overview page gives a model handoff that hits all five:

Pull this week's pipeline review list from our CRM. Skip anyone already in an active sequence. Research the top five accounts across the web and Slack, pull the right contacts, draft outreach in my voice, and leave me drafts to approve by tomorrow morning.

Outcome (outreach drafts), sources (CRM, web, Slack), constraint (skip active sequences), deliverable (drafts by tomorrow morning) and stop point (drafts to approve, nothing sent). The docs describe it as telling the Bot “what to do, where to work, what context to pull in, and what finished work looks like.”

One Bot, one job

The Bots guide recommends a separate Bot when the work has a distinct goal, set of tools and sources, working style, approval boundary or schedule. Jobs like “Talent Scout” or “Bug Reproduction” give a Bot guidance; “a General Helper does neither.”

Put the job in the Bot's first message and ask it to remember. The getting-started example:

Your job is product performance. Investigate product-performance questions with our observability tools. Preserve links and screenshots, separate evidence from hypotheses, and lead with the highest-impact issue. Never change production settings. Remember this for future work.

The official Grok Bot 101 guide adds a useful habit: describe the workflow “exactly as you would do it,” or record a demo of it.

A template you can copy

Job: You own [one outcome], for [who it's for].
Sources, in order of trust: [system 1], [system 2]. If a source is down, say so; don't guess.
Deliverable: [format], saved to [place], by [time].
Finished means: [a check someone else could do].
Stop and ask me before: sending, posting, paying, deleting, or changing [system].
Never: [the one thing that would be a disaster].
If there's nothing new: [say nothing / one line "nothing new"].
Remember these rules for future work.

Before and after

Before: “Keep an eye on our competitors and let me know what's going on.”

After: “Every Monday, check the pricing pages and changelogs of [A], [B] and [C]. List only changes since last Monday, each with a link and the date you saw it. Separate what changed from what you think it means. Post the list here. If nothing changed, reply with one line. Don't email or post anywhere else.”

The second one has a source list, a comparison period, a format, a quiet path and a boundary. It can also become a skill and then a routine without rewriting.

Ask for evidence, not just answers

xAI's tips for effective conversations: lead with the result you need, link the source of truth, state what the Bot may change and what needs approval, ask for evidence (links, screenshots or a short action log), and correct lasting preferences explicitly.

For anything consequential, also ask for a plan first. Flavio Copes suggests a five-line version in his guide: list the sources and tools, name every system you may change, define the deliverable, mark the steps that need approval, explain how you'll verify the result, then wait.

Why briefs fail

Memory helps, but it isn't a database. The docs advise keeping changing facts in the source system and asking the Bot to reopen current data for consequential decisions (Bots guide).