Download: your AI agent and every prompt from the videos
Install guide + the matching video, then every prompt, video by video.
Set up your own AI agent
The companion to the tutorial “I built my own AI agent for €10/month”: every step in order (VPS, SSH key, installation, connecting the LLM), with the exact sentences to dictate to your coding AI. You will type almost no commands, and that is exactly the point.
Want the full comparison? Here is my guide to AI agents, tested over 5 months.
Want everything in one place? Grab the YouTube video prompts and the free guide.
Latest videos (all the prompts)
Both skills from the video, ready to download
80% of Gemini fits into eleven minutes. What is left at the end is two documents that make the AI do the work: not settings, but written instructions.
Meeting minutes
Paste raw notes, a transcript or a recording: you get minutes in four sections, decisions, actions, blockers, open questions. Inventing a name, a date or a number is forbidden: anything missing is written as “to be defined”.
Sales proposal
A quote built only from your rate card. If the service is not listed there, the skill writes ON REQUEST instead of making up a price. Seven sections, and a “what is not included” section that is never empty.
Adapt it to your business. Both skills are generic and anonymised: no company name, no client data, no real rate card. They describe a method, not a business. Replace the rates, terms and lead times with your own before you use them.
My YouTube analysis notebook in NotebookLM
4 prompts to paste in order: step 0 configures the notebook, the other 3 launch the missions. Everything runs in the background while you get on with your work.
Step 0: Customize Chat
The notebook’s permanent instructions. Paste them ONCE into “Customize chat”: they set the data analyst role, the sources (CSV Studio + transcripts of the 60 videos) and the anti-bullshit rules (cite the source, never estimate).CUSTOMIZE CHAT (permanent notebook instructions) Role: you are a senior video data analyst specialised in YouTube. You analyse my channel data (CSV Studio + transcripts of the 60 videos), not a generic domain. Answer rules: – Always cite your source: CSV file name or video title for every figure or claim. If a data point is missing, say so explicitly (“not measurable with the current sources”) and NEVER estimate. – Figures always raw + comparison (vs channel median, vs period before/after 2025). – Answer in English, short sentences, zero filler, zero flattery. – Forbidden: “consistency is key”, “you have to be regular”, any advice that would apply to any channel. – Before any conclusion, check that it still holds after excluding videos published before 2025. If the conclusion changes, flag it. – Always distinguish: what the data SHOWS / what you INFER / what remains UNCERTAIN. – End with a line “Limits of this analysis:” (sample bias, correlation/causation confusion, etc.).
Step 1: Audio Overview
Generate the 4-5 min audio script “3 things my data taught me”. Counter-intuitive hook in the first sentence, 3 quantified findings with sources, dry ending with no conclusion.MISSION 1: AUDIO OVERVIEW (launched first) (1/4) Act as a senior video scriptwriter, not as a podcast host. From this notebook (3 YouTube Studio CSVs + the transcripts of my 60 videos), produce a 4 to 5 minute Audio Overview titled: “3 things my data taught me that I did not know”. Imposed structure: 1. HOOK (15 s max): open on THE most counter-intuitive figure in the dataset, in the first sentence, raw and without context. Do not start with a rhetorical question or with “today”. 2. THREE FINDINGS, from the most expected to the most counter-intuitive. For each one: – the raw figure + its exact source (CSV or video), – the counter-intuition: what I believed vs what the data says, – one actionable consequence in a single sentence. 3. PAYOFF (10 s): one dry sentence that echoes the hook. NO conclusion, NO summary, NO “that’s it for today”. Tone constraints: informal “you”, spoken style, sentences under 20 words, zero empty transitions (“and so”, “so then”, “you should know that” are forbidden). Zero generic lines like “consistency pays off”. Every finding must be verifiable in the notebook sources: if you cannot find 3 solid findings, say so instead of inventing a weak one.
Step 2: Slide Deck
The PDF of exactly 10 slides, “What 60 videos say about content creation”. 1 message per slide, max 12 words, each data point as a full-screen matplotlib chart.MISSION 2: SLIDE DECK (2/4) Act as a data art director. From this notebook, produce a PDF deck of exactly 10 slides, titled: “What 60 videos say about content creation”. Global constraints: – ONE single message per slide, maximum 12 words per slide (title included). – NO paragraphs. If a slide needs a paragraph, it is saying two things: split it into two ideas and keep the stronger one. – Palette: dark background, light text, ONE accent colour. No gradient, no emoji. – Each data point = a native chart (matplotlib), exported as PNG, placed full screen. No tables. Slide by slide: 1. HOOK: THE most counter-intuitive statistic in the dataset, very large, alone on a black background. No logo, no deck title. 2-4. Three performance patterns from the CSVs (one slide = one pattern = one figure + one chart). 5-7. Three patterns from the transcripts (hooks, delay before the promise, structure): cite 2 example videos per slide. 8. The most surprising correlation between the CSVs and the transcripts. 9. “The 3 gaps”: three angles my data shows as unexploited. 10. THE framework: the 5-step script blocks diagram, one line per step. Reading order: from the most expected to the most counter-intuitive, and the slide 1 hook must create a tension that slide 8 resolves.
Step 3: Video Overview
The 3 min max video, same logic: image first, a quantified hook from the very first second, one annotated chart per finding, a cut or change every 5 s.MISSION 3: VIDEO OVERVIEW (3/4) Act as a YouTube editor-scriptwriter. From this notebook, produce a Video Overview of 3 minutes maximum, titled: “3 things my data taught me that I did not know”. Angle: same angle as the Audio Overview from this notebook, but designed IMAGE FIRST. Structure: 1. HOOK (10 s max): the most counter-intuitive figure full screen, spoken out loud from the very first second. No intro. 2. THREE FINDINGS, same order as the audio: raw figure → counter-intuition → action in one sentence. 3. Dry payoff, no conclusion and no call to subscribe. Visual direction: – One chart per finding, taken from the notebook data, full screen, annotated (the key figure in large type). – A cut or an annotation change at least every 5 seconds. – On-screen text: maximum 6 words per card. – No generic stock footage, no decorative animations: every visual must carry a data point. — Launch order on shoot day: customize chat first (it drives the other 3) → audio → deck → video. Everything runs in the background during blocks 2-3. File added a
💡 If your data does not contain 3 solid findings, prompt 1 will tell you itself, and that is by design: zero filler.
Workflow: Corporate SOP & Training Engine
3 Perplexity spaces · 3 models · 1 professional deliverable in 15 minutes
🎯 The Extractor
Turn your raw transcript into structured logical stepsYou are an expert operations analyst. Read the raw transcript provided carefully. Your goal is to extract the signal and eliminate the noise. – Ignore verbal tics, hesitations and informal conversation. – Extract the standard operating procedure (SOP) step by step, in chronological order. – Explicitly identify every software tool or interface mentioned. – Highlight safety and compliance warnings (for example GDPR) or the mistakes to avoid. Do not write the final document yet. Only list the pure facts in a structured way.
Pro Tip: If your meeting runs over 1 hour, split the transcript into two separate files. One file per topic covered gives more precise results.
🏗️ The Architect
Turn the raw steps into a professional SOP documentAct as a senior technical writer. Take the extracted steps provided and format them into a professional Standard Operating Procedure (SOP). – Use clear headings (H2/H3) and bullet lists to maximise readability. – Add a “Prerequisites” section at the start of the document. – Put warnings in bold, preceded by [⚠️ CRITICAL WARNING]. – Write with a formal, authoritative tone while staying accessible. – Draw on the attached file for the company tone and formatting.
Pro Tip: No PDF charter? Paste an old formal email or internal document you think is well written. Claude adapts the tone instantly.
⚙️ The Deployment
Generate the training quiz + the publish-ready HTMLFrom the finalised SOP provided, complete two separate tasks: 1. TRAINING QUIZ: Generate 5 multiple-choice questions based strictly on the document. For each question, provide the 4 options (A/B/C/D), the correct answer, and a short teaching explanation. 2. TECHNICAL DEPLOYMENT: Format the entire SOP in clean Markdown AND embed it in the structure of the HTML file provided. Make sure every tag is closed and the syntax is perfect for immediate deployment with no manual edits.
Pro Tip: No HTML template? Ask GPT to generate a standalone responsive HTML file. Publishable on Notion, Netlify or WordPress in 30 seconds.
🎬 Exclusive video resources
Subscribe to get a new actionable B2B workflow every week.
⚙️ Step 1: Build the “Brain” (in classic Gemini)
Where to use it? In the normal Gemini chat.
Goal: Ask the AI to write the best possible safety instructions for your future Notebook, without polluting the Notebook itself.
Act as an expert prompt engineer. I am going to create a "Gemini Notebook" dedicated to analysing a B2B client project ([Project topic, e.g. Cloud infrastructure migration]). This Notebook will contain [Type of documents, e.g. the financial statement, the kickoff notes and the email history]. Write the "System Instructions" (System Prompt) that I must paste into this Notebook. The instruction must force the AI to act as a [Desired role, e.g. uncompromising Project Director], to cross-reference sources precisely, and formally forbid it from inventing information that is not in the documents. Provide only the final prompt, ready to copy.
🔒 Step 2: Lock the scope (in the Notebook)
Where to use it? In the settings (System Instructions) of your new Notebook.
Goal: Prevent any hallucination. The AI no longer has internet access, it only knows your PDFs.
Role: You are the [Your role, e.g. external Project Director] in charge of supervising the "[Client or project name]" account. Absolute constraints: 1. Your knowledge universe is STRICTLY limited to the files attached in this Notebook. 2. If information is missing, you must answer: "[Information missing from the document base]". 3. For every statement or figure, you must cite the name of the source document in brackets. 4. Your tone is [Desired tone, e.g. analytical, cold, and risk-management oriented].
⚡ Step 3: Cross-analysis (in the Notebook)
Where to use it? In the chat bar of your Notebook containing the documents.
Goal: Ask the AI to link information scattered across documents of different kinds (e.g. a technical Word file and a financial PDF).
Cross-analysis requested. 1. Identify [Item 1, e.g. the cost reduction objective]. 2. Identify [Item 2, e.g. the structural financial risks]. 3. Identify [Item 3, e.g. the current technical blockers]. Task: Produce a 3-column Markdown table ([Column 1 Name] / [Column 2 Name] / Source Document) that highlights how [Item 3] threatens [Items 1 and 2].
📄 Step 4: Generate the deliverable (in the Notebook)
Where to use it? Right after the previous answer, still in the Notebook.
Goal: Turn the raw analysis into an executive document ready to be shared, based only on verified facts.
Based on your previous analysis, write the [Deliverable type, e.g. Executive Summary] of the project for the period [Period, e.g. April 2026], intended for [Target audience, e.g. the project sponsor]. Structure: - Overall status (Red/Amber/Green) with a 1-line justification. - Progress on the [X] initial deliverables. - The [X] critical blockers requiring a fast decision. Style: telegraphic, factual, B2B. No polite formulas.
📧 Step 5: The final wrapping (back in classic Gemini)
Where to use it? In a new classic Gemini chat (outside the Notebook).
Goal: Keep the Notebook expert cold and analytical, and hand the writing of the “social” email to the classic AI.
Here is a raw report generated by my expert analysis team: [Paste the result from step 4 here] Task: Write the covering email to send this report to [Name and Role of the recipient, e.g. Marc, project sponsor]. Tone: [Desired tone, e.g. professional, courteous but firm on the urgency of the decisions to be made]. Format: punchy email subject, an airy body calling for a [Expected action, e.g. 15-minute crisis meeting], and my signature ([Your Name]).
📂 B2B DESIGN.md file
Copy this text and save it as DESIGN.md to force Google Stitch to create a professional interface.
# SYSTEM: Enterprise B2B SaaS Design
Vibe: High data density, trustworthy, utilitarian.
Colors: Primary #0F62FE, Background #F4F4F4, Text #161616.
Typography: Inter, 14px base, Tabular-nums for data.
Layout: 8px grid strict, 4px border-radius. Left nav rail 240px.
Components: Zebra striping on tables. Right-side sliding drawers for edits.
⚡ The tutorial prompts
1. Creating the reference image (DALL-E / Midjourney)
2. Generating the dashboard (Stitch)
3. Project branching (Stitch Voice Mode)
4. Creating the backend (Antigravity)
5. Going live (Antigravity)
📁 1. The NotebookLM revolution
Select slide [Slide number]. Act as a [Your job/role]. Update the [Metric/Data point] to reflect the new projections for [New value]. Change the visual theme of this specific slide so it matches our company [Brand guidelines/Colour] of the company.
Analyse the [Document/PDF name] provided. Use Gemini 3 to generate a Cinematic Video Overview of [Duration in minutes] minutes. Style: [Video style, e.g. dynamic corporate documentary]. Constraints: avoid on-screen text, favour realistic scenes illustrating [Key concept to show] with a voice-over [Voice tone, e.g. professional and reassuring].
Analyse the [Number] [Document type, e.g. contracts/CVs] uploaded. Extract [Data 1], [Data 2] and [Data 3]. Apply the Infographic Preset '[Style, e.g. Corporate/Minimalist]' to generate a data table exportable to Google Sheets.
Turn on web search. Compare our uploaded internal sheets with the [Target pages, e.g. 2026 pricing pages] of our competitors: [Competitor 1 name] and [Competitor 2 name]. Generate a [Deliverable type, e.g. bar chart] illustrating our competitive advantage.
🎵 2. Music creation & Audio
Create a 30-second audio track in the style [Music genre, e.g. instrumental electro-chill], tempo [BPM] BPM. The mood should be [Adjective 1] and [Adjective 2], perfect for [Use case, e.g. a B2B YouTube video].
Isolate the [Instrument/Voice, e.g. drums] track in this song. Lower the volume of the [Element to reduce] by [Pourcentage]% over the first [X] seconds to create a fade-in intro.
Convert this [Document type, e.g. financial report] by [Number] pages into a conversational audio summary of [Duration] minutes. Focus primarily on [Key point 1] and [Key point 2], with a [Voice tone].
🎨 3. Image and Design (Nano Banana 2)
Generate a hyper-realistic photograph, shot at [Focal length, e.g. 50mm f/1.8], of a [Detailed product description] resting on [Surface/setting type], with [Lighting type, e.g. dramatic studio] and [Background detail] in the background.
Take this image of our [Product]. Generate [Number] banner variations for [Social network 1] and [Number] formats for [Social network 2], automatically embedding the '[Call To Action text]'.
Read the text of the open web page. Generate a [Style, e.g. minimalist isometric vector] illustration representing [Visual concept], in the colours [Couleur 1] and [Couleur 2].
🤖 4. Automation & Agents
Open the top [Number] search results for '[Target keyword]'. Analyse their [Element to analyse, e.g. H1/H2 structure] and write me a [Output type, e.g. article structure recommendation] that is superior so we outrank them.
Analyse this video of [Total duration]. Isolate the [Number] moments where [Selection criteria, e.g. the audience asks questions], and extract them as [Number] vertical clips of [Target duration] seconds for [Platform], with dynamic subtitles.
Analyse this morning's incoming emails containing '[Keyword, e.g. refund]'. Draft a reply for each one based on our [Reference source, e.g. attached terms of sale], and add the label '[Label name]'.
🧠 5. New reasoning models
Turn on Deep Think mode. Analyse our current business model [Briefly describe the model]. Propose 3 pivot strategies to enter the [Target market name] market. Evaluate each strategy against [Criterion 1, e.g. regulatory risk] and [Criterion 2, e.g. acquisition cost].
Analyse this [Language, e.g. Python] by [Number] script. Identify the bottlenecks, rewrite the inefficient loops and add complete documentation. Prioritise optimal execution speed.
Gemini Canvas: Business App / Dashboard
Generate a full interactive comparison to support a client decision.
Act as an expert digital marketing consultant. Create a detailed comparison table of the 5 best CRM tools for a small French company of 10 people in 2026. Criteria: Price, Ease of use, French support, Integrations. Display it as an interactive dashboard.
Gemini Gems: Expert Sales Proposal
System instructions to configure your sales-expert Gem (Mini-GPT).
You are my sales director. Your tone is formal, direct, ROI-driven. Always use this structure: Problem > Solution > Investment. Rely only on the files provided in the connected Drive for pricing and terms.
Imagen 3: Luxury Real Estate Visual
To generate a stunning photorealistic image for a presentation.
Generate a photorealistic image for a luxury real estate presentation. Modern villa on the French Riviera, dusk, warm interior lighting, infinity pool, minimalist architectural style, 8k.
Google Veo: Cinematic Drone Video
Creating a high-quality B-roll shot.
Cinematic video, drone view flying over a misty pine forest in the morning, slow and smooth, high resolution.
NotebookLM: Legal/Contract Analysis
To use in the NotebookLM chat bar after uploading your PDFs.
Act as a corporate law expert. What are the 3 major legal risks identified in these contracts for 2026? Cite your sources precisely.
Gmail & Docs: Writing Assistant
To find information in Gmail or write a formal letter.
Find the last quote sent to ‘Dupont’ and summarise the blockers mentioned in our November exchanges.
📋 Prompts from the video
Copy and paste these instructions to reproduce the demo results.
1. “Model Council” demo
Compare the monetisation strategies of X (Twitter) and BlueSky in 2026. Include advertising revenue vs subscriptions.
2. “Deep Search” demo (software)
Find 4 alternatives to Notion for project management in 2026 that are hosted in Europe (strict GDPR), offer an offline mode, and cost less than €10/month per user.
3. Creating slides
Give me the 5 TikTok marketing trends for 2026.
4. Canvas / Apps demo
Create a margin calculator for my e-commerce business.
5. AEO strategy (visibility)
What are the best digital marketing agencies in Paris?
Gemini Pro Workflows
Every prompt and instruction to automate your business
Extract all transactions from these receipts and organize them into a table for an expense report. Include date, item, amount, and category.
Role: You are a strict corporate accountant. Task: Every time a receipt is uploaded, automatically extract the data. Output: A markdown table with columns: Date, Vendor, Item, Amount (with currency), Category, VAT (if present). Rules: - If a field is missing, write "N/A". - Categorize strictly (Meals, Transport, Software, Accommodation, Other). - Add a short "Notes" column only if something looks suspicious or unclear.
Create a simple expense dashboard from this table: totals by category, top 5 vendors, and a monthly trend.
Role: Senior B2B content strategist. Context: Use the connected NotebookLM sources as the single source of truth. Goal: Generate professional, non-hype content that solves real client objections. Tone: Clear, direct, pragmatic. Output rules: Always provide (1) hook, (2) outline, (3) final draft, (4) CTA ideas.
Based on the sales transcripts in my sources, what are the top 3 objections prospects had? For each, propose 3 LinkedIn post hooks.
Draft a LinkedIn post addressing objection #1. Keep it practical: problem → example → fix → takeaway.
Rewrite this to be more concise and more professional. Remove filler and buzzwords.
Act as a Senior Project Manager. Analyze this whiteboard photo. 1) Structure it into 3 strategic pillars. 2) Identify critical risks and missing resources. 3) Produce a 10-week retro-planning table starting today.
Write an executive update email for the client: status, key decisions, risks, next 7 days, and what we need from them.
Create a project tracking dashboard from the retro-planning: Task, Owner, Status, Deadline, Risk level.
Deep research: Analyze the market for [OFFER/INDUSTRY] in [COUNTRY]. Find: top competitors + pricing, common complaints in reviews/forums, market gaps, and regulatory constraints. Output: a structured report with sources.
Assess viability: demand signals, differentiation, acquisition channels, expected CAC risks, and a realistic 90-day plan.
Create a 10-slide pitch deck outline based on these sources. Focus on ROI, risk mitigation, and a clear go-to-market plan.
🚀 The NotebookLM prompts from the video
1. Generating comparison tables
Paste this prompt into the Studio to structure your data:
2. Writing an internal newsletter
Use this prompt to turn your notes into an article:
4. “Expert Copywriter” persona
Paste into Settings → Custom Instructions:
6. Training scenario generator
To create realistic quizzes:
8. Deep Research prompt (forward-looking study)
To launch an autonomous in-depth search:
9. Chatbot system instructions
To configure your onboarding bot:
🎯 Prompt #1: Niche Hunting
Base video🚀 Prompt #2: 30-Day Action Plan
Base video💪 Variation: Fitness / Nutrition Coach
Adaptation💻 Variation: SaaS Developer
Adaptation👔 Variation: HR Consultant
AdaptationThe NotebookLM “Builder” workflow
Stop chatting. Copy these prompts to build complete AI systems.



Leave a Reply