Weak prompts waste re-open time. These before-and-after cards show a tighter Grok Bot job you can copy. Workshop patterns — not official xAI docs. Examples are illustrative.
Watch these sites and tell me if anything important changes.
Better version — Compare-sources routine
Compare only the sources I name. Do not watch the open web.
Sources (only these):
- {{source_a}}
- {{source_b}}
Cadence: {{cadence|weekly}}
Decision this comparison supports: {{decision}}
Deliver:
1) Table: Topic | Source A says | Source B says | Match?
2) Conflicts (quote lightly)
3) Gaps in either source
4) What changed since the last brief, if I attached one
5) Recommendation options — do not decide for me
Done when the named topics are covered and every claim has a source URL.
Do not email, post, or overwrite files.
Why it wins
A vague watcher produces noise or silence because “important” is undefined. A compare-sources routine names the pages, the cadence, and the decision, so you get a reusable table. Pair with the Compare documents workflow. Examples are illustrative.
You are researching for me. Stay in gather-and-summarize mode.
Goal: Produce one research brief on: {{topic}}
Audience: {{audience|general}}
Depth: {{depth|standard}}
Rules:
- Prefer primary / official sources when making factual claims.
- Separate Facts / Unknowns / Opinions.
- List sources with URLs.
- Do not send email, post publicly, or edit live systems.
- Stop and ask before any action that leaves this chat.
Deliver:
1) 5–8 bullet findings
2) Source list
3) Open questions
4) Suggested next step (draft only)
Why it wins
Ten tabs and half-remembered claims waste re-open time and invite wrong citations. One brief with Facts / Unknowns / Sources is a single artifact you can reopen. Pair with the Research brief workflow. Examples are illustrative.
Keep working on this until it’s good. Just send a reply if you need to.
Better version — Done criteria
Draft a customer reply. Stop when the draft is ready for my review.
Channel: {{channel|email}}
Tone: {{tone|friendly-professional}}
Customer wrote:
{{customer_message}}
Trusted facts (only use these; if missing, ask):
{{facts}}
Deliver:
1) Reply draft
2) Assumptions I made
3) Risks if sent as-is
4) Optional shorter variant
Done criteria:
- No send, refund, credit, or calendar hold
- Every commitment is marked draft
- Unknowns are listed, not guessed
Hard stop: do not send, post, or update the CRM.
Why it wins
“Until it’s good” never ends, and replies from memory over-promise. Done criteria make the stopping point reviewable: a draft plus Assumptions and Risks, nothing sent. Pair with Draft a customer response. Examples are illustrative.
Share these notes with the team and make sure someone follows up.
Better version — Single owner
Turn these notes into clean minutes and actions.
Timezone: {{tz|America/Chicago}}
Attendees: {{attendees}}
Raw notes:
{{notes}}
Deliver:
1) Summary (5 bullets max)
2) Decisions
3) Action table: Action | Owner | Due | Notes
4) Unresolved questions
Rules:
- One owner per action. Do not invent owners or deadlines — mark Unknown.
- Do not email attendees or create calendar events without approval.
- Flag speculative items separately.
Why it wins
Notes that sit, or that belong to “the team,” stall in follow-up pings. An action table with Unknown owners flagged gives one owner per row — and nothing is emailed until you approve. Pair with Meeting notes → actions. Examples are illustrative.
Here’s the whole process again. Do this every week the same way I just described.
Better version — Skill then routine
Write the method once, then reuse it on a cadence. Do not rebuild the instructions each run.
Skill (the method — write once):
- Job: {{job}}
- Required inputs
- Steps
- Done criteria
- Approval boundaries (no send / post / purchase)
Routine (the cadence):
- Frequency: {{cadence}}
- Active hours: {{hours}}
- Output: the same finished shape each run
Do not start the routine until I approve the skill text.
Do not invent product limits or installer URLs.
When stating Grok Bot product behavior, verify against https://docs.x.ai/grok-bot/
Why it wins
Rewriting the same prompt each week burns time and drifts. A skill holds the method; a routine holds the schedule. Official Grok Bot docs describe saving a demonstrated flow and reusing it later, including on a schedule — verify current product behavior there. Use the routine frequency calculator to count scheduled runs, not spend. Examples are illustrative.
