2025-06-20 Initech Riga Seminar on AI

2025-06-20 Initech Riga Seminar on AI1. Materials / Video2. Prompt engineering basicsStructure of prompt RulesRecommended first words to use3. Internet search4. RAG (Retrieval Augemented Generation)5. Checking Criteria and Guidelines 6. Making MermaidJS charts7. Working with Excel

1. Materials / Video

Figmaboard: https://www.figma.com/board/t2zgLDvURk4x4qoF8frcgi/2025-06-20-initech.no-Seminar-GenAI?node-id=0-1&t=eJql6uin6MSGXhdd-1

📺 Video from Session #1 https://youtu.be/7trVSPMieNk

📺 Video from Session #2 https://youtu.be/qIFPaEn0XoY

Presentation: https://share.yellowrobot.xyz/quick/2025-6-23-6FE3041A-1058-40FE-A1C5-3F75353C20C0.pdf

2. Prompt engineering basics

Use as little words as possible for Goal/Task, Return sections, but use extensive facts in “Context Dump” part

Structure of prompt

Rules

Write each topic in a new session
Do not continue chats except if you have formating tasks, instead press edit of previous chat message
Write tasks and keywords in English
Begin the task section with a verb
Write terms and keywords with Capital Initial Letters, combine terms with quotation marks
Correct grammatical errors
Structure the request using separating symbols ``` ... ``` , ---
Add as many Facts and Examples as possible (x10 more important than anything else)
Output / Result formatting
Acknowledge limitations (4-6 pages of text)
Do not use files, instead copy as text in Contect Dump section
If you do not like result run again, because it is always little bit random and next try might give better result

Recommended first words to use

create list
rewrite sentence
rewrite paragraph
generate list
generate sentence
write sentence
write story
explain concept
explain difference
give example
edit sentence
describe process
describe difference
describe impact
classify sentence
classify item
classify concept
summarize article
summarize text
find word
find synonym
translate

Dieagram of first words in common LLM datasets.

Prompt examples:

🤖 Use o3-pro or o3 model without Internet Search (make sure you disable it).

Linkedin post


x
1
Write LinkedIn post based on the text as a critical post objecting to the article.•Output: In English, Business Style, 1-2 Sentences with Numbers and Facts.
2

3
Post:
4
```
5
{copy_post}
6
```

Summary of Article


xxxxxxxxxx
1
1
Write summary of Article.
2
Output 1-3 sentences, include Facts, in Norwegian Language.
3

4
Article:
5
```
6
{copy_text}
7
```

Translate


xxxxxxxxxx
11
1
Translate Text Norwegian to English language.
2

3
Examples of translations:
4
```
5
{optional_examples_of_translation}
6
```
7

8
Text:
9
```
10
{copy_text}
11
```

Challange to invoice


xxxxxxxxxx
1
10
1
Write challenge to invoice about software development services so I would not need to pay it. •
2
Output: In English, as email without sections, formal style, ready to send•
3
Warning: Leave no templates.
4

5
Invoice:
6
```
7
{copy_text}
8
```
9

10
Agreement:•
11
```
12
{copy_text}••
13
```
14

15
Company Polcies:•
16
```
17
{copy_text}
18
```

3. Internet search

For quick internet search adice to use Pro version of:

For precise in-depth search use following

Inside ChatGPT o3

Instead of asking directly to compare or qualify facts on internet split task in 2 phases:

Phase 1: Information gatgering

🤖 Use o3-pro or o3 model with “Research” button enabled.

Leave Resarch chats working in background while doing other things


xxxxxxxxxx
1
1
Find Wind Electricity production for Norway 2023, Include MW Capacity, Production TWh.

Then


xxxxxxxxxx
1
1
Find Wind Electricity production for Sweden 2023, Include MW Capacity, Production TWh.

Phase 2: Information processing

🤖 Use o3-pro or o3 model without “Research” (make sure Internet search is turned off).


x
11
1
Compare Wind energy produced in Norway, Sweden in 2023. Which Policies overlap, which do not overlap.
2

3
Norway:
4
```
5
{copy_norway_research_results}
6
```
7

8
Sweden:
9
```
10
{copy_sweden_research_results}
11
```

Finally you could ask to produce table in same chat using GPT4.5


xxxxxxxxxx
1
1
Make Table with columns "Country", "MW, Capacity", "Production TWh"

🧩 Example how to gather and display infomration about GJF:OSLO stock https://chatgpt.com/share/685944fe-629c-800a-8890-6a60b3c676fe

4. RAG (Retrieval Augemented Generation)

To find quickly something by semantic meaning in your own Documents or Project proposals etc. You would use RAG, it can be used either by adding files to ChatGPT or ChatGPT Projects or Google NotebookLM which is currently the best solution.

In ChatGPT projects use this functionality like this:

In Google NotebookLM use this functionality like this: (Also recommened to use Google NotebookLM app) https://notebooklm.google.com/

5. Checking Criteria and Guidelines

To check if a document matches guidelines or criteria, do not use 2 documents in the prompt. Instead, split the task into 2 phases and execute it as a textual task.

Phase 1: Prompt using o3-pro or o3

Convert text guidlines into yes/no questions.


xxxxxxxxxx
1
1
Write All Criteria for contractor in English based on Requirements. Criteria should be Yes / No questions. Before for each Criteria quote Requirements as reasoning.
2

3
Requirements:
4
```
5
{copy_requirements_or_guidlines}
6
```

Followup prompt to re-format text using GPT 4.5


xxxxxxxxxx
1
1
Extract Questions as numbered list. Don't change questions.

Phase 2: Prompt using o3-pro or o3

Test documents using Yes/ No questions.

https://chatgpt.com/share/68594784-8f68-800a-a012-e8d8df74aa3c

For reasoning tasks, it is a good idea to ask to quote reasoning before each answer because LLMs tend to copy output from input, and providing a correct quote near the final answer should help in producing better answers.


xxxxxxxxxx
3
10
1
Answer each Criteria using Yes / No for Document as numbered list. BEFORE each Criteria write Quote text from Document for supporting Answer. Bold those that answer is No.
2

3
Criteria:
4
```
5
{copy_yes_no_criteria}
6
```
7

8
Document:
9
```
10
{copy_document}
11
```

https://chatgpt.com/share/685946b9-b154-800a-af7c-eaee5ca01029

6. Making MermaidJS charts

MermaidJS format documentation: https://mermaid.js.org/intro/

Prompt (o3-pro):


xxxxxxxxxx
1
1
Make MermaidJS flowchart diagram of Process. Use Decision nodes with Yes/No. Make it in English. Don’t use HTML tags in node titles. Don’t include any special symbols in node titles.
2

3
Process:
4
```
5
{copy_text_of_document}
6
```

https://chatgpt.com/share/68594541-49d4-800a-9f48-52c003c441e7

For display and further editing using mouse and AI use: https://www.mermaidchart.com/

7. Working with Excel

Must have a clear CSV format without merged columns. Each Excel sheet is a separate CSV file.

Prompt (important to use GPT4.5 not o3):

Upload CSV
Apply mofifications and tasks one by one


xxxxxxxxxx
1
1
Calculate average sales by "Customer type"

Visualize


x
1
Plot Pie Chart, total revenue For "Product Line" where Total is above 500

https://chatgpt.com/share/68594668-bf38-800a-8d93-26f8b2556f5f