[Minimalist food photograph, [1080x1080] – a single [OBJECT] rests on a light, matte surface and is captured mid-transformation into a lego bricks 3D form: one half remains intact while the other organically fragments into large, floating lego bricks that drift outward, each brick revealing the object’s texture, ingredients, and colors. Studio lighting with soft, realistic shadows, shallow depth of field, tasteful perspective and composition, hyperrealistic detail, stylish geometric abstraction, high resolution, cinematic close-up]
The Core: In this tutorial, you will learn how to create a CRM dashboard for your sales pipeline with Notion AI. Instead of manually creating pages, subpages, and toggles, all of this will be done for you using AI prompts.
Step-by-step:
- Go to Notion.so, open Notion AI, and collapse the side panel for a cleaner workspace
- Use this prompt:
- Click Submit and let Notion AI build the database structure, then agree to suggested additions like follow-up stages and next steps
- Review the generated "CRM – Sales Pipeline" page and refine fields by changing toggles to checkboxes or adding filters like deal size
Step-by-step:
- Go to Notion.so, open Notion AI, and collapse the side panel for a cleaner workspace
- Use this prompt:
"A CRM pipeline database where I can track leads from first contact through conversion. Add filters for email, type of subscriber, first contact, second contact, did they convert, and why didn't they convert"- Click Submit and let Notion AI build the database structure, then agree to suggested additions like follow-up stages and next steps
- Review the generated "CRM – Sales Pipeline" page and refine fields by changing toggles to checkboxes or adding filters like deal size
Pro Tip: Experiment with prompt styles. For example, instead of saying "create me a CRM," try "create a sales pipeline with Kanban stages" or "build a lead tracker with follow-up reminders." The more specific you are, the better your results.
Spent 6 months deep in #prompt engineering. Here's what actually moves the needle:
Getting straight to the point:
1. Examples beat instructions Wasted weeks writing perfect instructions. Then tried 3-4 examples and got instant results. Models pattern-match better than they follow rules (except reasoning models like o1)
2. Version control your prompts like code One word change broke our entire system. Now I git commit prompts, run regression tests, track performance metrics. Treat prompts as production code
3. Test coverage matters more than prompt quality Built a test suite with 100+ edge cases. Found my "perfect" prompt failed 30% of the time. Now use automated evaluation with human-in-the-loop validation
4. Domain expertise > prompt tricks Your medical AI needs doctors writing prompts, not engineers. Subject matter experts catch nuances that destroy generic prompts
5. Temperature tuning is underrated Everyone obsesses over prompts. Meanwhile adjusting temperature from 0.7 to 0.3 fixed our consistency issues instantly
6. Model-specific optimization required GPT-4o prompt ≠ Claude prompt ≠ Llama prompt. Each model has quirks. What makes GPT sing makes Claude hallucinate
7. Chain-of-thought isn't always better Complex reasoning chains often perform worse than direct instructions. Start simple, add complexity only when metrics improve
8. Use AI to write prompts for AI Meta but effective: Claude writes better Claude prompts than I do. Let models optimize their own instructions
9. System prompts are your foundation 90% of issues come from weak system prompts. Nail this before touching user prompts
10. Prompt injection defense from day one Every production prompt needs injection testing. One clever user input shouldn't break your entire system
The biggest revelation: prompt engineering isn't about crafting perfect prompts. It's systems engineering that happens to use LLMs
Hope this helps
Getting straight to the point:
1. Examples beat instructions Wasted weeks writing perfect instructions. Then tried 3-4 examples and got instant results. Models pattern-match better than they follow rules (except reasoning models like o1)
2. Version control your prompts like code One word change broke our entire system. Now I git commit prompts, run regression tests, track performance metrics. Treat prompts as production code
3. Test coverage matters more than prompt quality Built a test suite with 100+ edge cases. Found my "perfect" prompt failed 30% of the time. Now use automated evaluation with human-in-the-loop validation
4. Domain expertise > prompt tricks Your medical AI needs doctors writing prompts, not engineers. Subject matter experts catch nuances that destroy generic prompts
5. Temperature tuning is underrated Everyone obsesses over prompts. Meanwhile adjusting temperature from 0.7 to 0.3 fixed our consistency issues instantly
6. Model-specific optimization required GPT-4o prompt ≠ Claude prompt ≠ Llama prompt. Each model has quirks. What makes GPT sing makes Claude hallucinate
7. Chain-of-thought isn't always better Complex reasoning chains often perform worse than direct instructions. Start simple, add complexity only when metrics improve
8. Use AI to write prompts for AI Meta but effective: Claude writes better Claude prompts than I do. Let models optimize their own instructions
9. System prompts are your foundation 90% of issues come from weak system prompts. Nail this before touching user prompts
10. Prompt injection defense from day one Every production prompt needs injection testing. One clever user input shouldn't break your entire system
The biggest revelation: prompt engineering isn't about crafting perfect prompts. It's systems engineering that happens to use LLMs
Hope this helps