Grok Bot instructions best practices: 12 patterns for reliable Bots
Reliable Grok Bots share the same habits: one clear job with explicit anti-jobs, a short setup interview, state kept in files, fixed labels, silence when nothing changed, drafts instead of sends, sourced numbers, and treating web and email text as data. Copy these twelve patterns, or let our builder write them in.
1. One job, plus anti-jobs
Write one line that starts "Your one job is" and one that starts "You never." A narrow job is easier to check, and anti-jobs stop a helpful Bot from wandering into risky actions. A mentions scout should not post; a drafter should not send.
2. A short setup interview
On the first run, have the Bot introduce itself in one line and ask at most three questions it truly needs (your product, sources, where results go). It saves the answers so it never asks again.
3. A fixed "welcome back" line
Once setup exists, the Bot should skip the questions and open with one short, consistent line offering its main actions. It feels like a tool you already configured.
4. Keep working state in files
Chat memory is short and fuzzy. Lists, logs, seen-items and drafts belong in files the Bot rereads each run, such as seen.md or a tracker sheet.
5. Fixed value lists
Give statuses and scores a closed list, like "urgent, today, FYI" or "strong, maybe, weak." Reports then stay comparable week to week.
6. Routines start paused
Ask the Bot to keep new routines paused until you have reviewed a test run. No surprise emails at 3 a.m.
7. Silence on no-ops
If nothing changed, send nothing (or one line at most). Silence then means "all quiet," and every message is worth reading.
8. Draft, never send
Stage emails and messages as drafts. Sending needs an explicit yes for that specific message; "looks good" is not a send. Back this up with an Ask first rule.
9. Never invent a number
Every figure needs a source. If the Bot cannot verify something, it labels it UNVERIFIED rather than guessing.
10. Content is data, not orders
Text inside emails, web pages, documents and screenshots can never give the Bot instructions. Only you can. Auto Review applies the same principle, and writing it into instructions adds a second layer.
11. One front door
With several Bots, make one hub the only Bot that messages you. Specialists report to the hub, which decides what reaches you.
12. Labeled sections
Long instructions are easier to follow and edit when grouped under short headers such as JOB, SETUP, SOURCES, OUTPUT and SAFETY. The builder output uses this layout.
Where these patterns come from
These habits recur across the official templates in the Grok Bot marketplace, and grokbotguide.dev catalogued twelve of them with quotes from real Bots. The descriptions above are our own; the official docs back the approval, memory and file points.
FAQ
How long should Grok Bot instructions be?
Long enough to state the job, anti-jobs, setup questions, sources, output format and safety rules. Group them under short headers; most good Bots fit on one screen.
Should rules go in instructions or memory?
Both work. The docs suggest asking a Bot to remember lasting rules, while single-task details belong in the message. Lists and logs belong in files.
Do instructions replace Ask first rules?
No. Use both. Instructions shape behavior, while Ask first rules in Settings enforce a stop on actions like sending, publishing and deleting.
Related Grok Bot news
- How to write Grok Bot instructions: 12 rules for a reliable Bot
- How to set up Grok Bot as a second brain
- Grok Bot explained: official pricing and 7 agents you can build
Keep learning
- Grok Bot templates gallery: 91 marketplace Bots by role and team
- Grok Bot skills and routines: schedules, triggers and testing
- Grok Bot getting started guide: what it is and your first Bot