Teach Grok Bot why you make exceptions
Log each recommendation, your final call, and your reasoning so future recommendations use past decisions.
What you’ll produce
A structured decision log (situation, recommendation, final decision, change, reason) that Grok Bot consults before similar recommendations.
Inputs & connections
- One workflow to start with, such as customer support, supplier negotiations, or sales
- A persistent store such as a Notion database
- Your written business policies
Requires separately configured Grok Bot access to the named data sources. Copying this prompt does not connect any accounts.
Your starting prompt
Replace the bracketed fields. This is our suggested prompt, not a quote from the source.
I want you to build a decision-learning system for the business decisions you help me make. Whenever you recommend an action and I approve, reject or modify it, capture: 1) the situation and relevant facts, 2) your original recommendation, 3) my final decision, 4) what I changed, 5) why I made that decision. If my reasoning is unclear, ask me one short question. Store these decisions in a structured log in [Notion database] that you can reference in the future. When a similar situation arises, retrieve relevant past decisions and explain how they apply before making your recommendation. Keep business policies separate from individual exceptions. Never assume one exception becomes a permanent rule. Start by helping me set up the decision log and choosing one business workflow to test.
How to run it
- 01
Connect a persistent database such as Notion.
- 02
Pick one workflow to test.
- 03
Make sure each entry captures your reasoning, not just approve or reject.
- 04
Review the log periodically to fix bad assumptions and remove outdated entries.
Review before you act
Keep human approval for consequential decisions and check that cited past decisions really match the new situation.
Limitations & success checks
- This does not retrain the model; it only gives Grok Bot access to past decisions.
- The author describes a proposed setup, not measured results; exceptions can be over-applied if the log is not reviewed.
Measure: Recommendations you accept without changes, and exceptions correctly flagged as exceptions.
Sources & attribution
The sources below support the reported example or a tool capability. They do not validate our prompt or prove the outcome in your environment.
Post by @Michael_Fenech_ (2026-10-10) with the full prompt and five setup recommendations, using a refund exception for a $15,000 customer as the example. The prompt is the author's suggested setup; no results are reported.
Creator / publisher: @Michael_Fenech_
Editorial review: 2026-10-14. How we label evidence.