The toolbox
My real stack: the tools I use every day and that make me massively productive, plus everything shown during the talk. Some links are referral links, they support the free content.

Claude Code on a computer that runs 24/7, available everywhere, connected to everything. It runs my business without limits.
I speak, the AI writes. A phenomenal time-saver.
Every multimedia model (video, image, voice…) in one place. No subscription, pay per use. My n°1 platform.
My n°1 choice to host my sites, apps, AI agents and my AIOS. Dead simple to get started.

Essential for all app development, when my AIOS is too busy.
Essential for client work: stay 100% focused without losing a crumb of any meeting. Mostly free.
Clinically proven AI focus music. +119% beta waves. Essential for deep work.

My brain for documents: I feed it sources, it digests them, I query it. Largely free, underrated.
N°1 for writing and code. Opus = one of the most powerful models available.
The best trained for natural communication. GPT Store, persistent memory.
Web-augmented search. Multi-model.
AI aggregator running on Google models. Pay per use.
The real gateway to Gemini. Far superior to plain Gemini.
Luxury AI. Autonomous multi-agent (acquired by Meta).
Prefer AI Studio to get the most out of it.
Solid on real-time news (X data).
The Chinese bomb. Not that incredible vs the competition.
The French champion. GDPR advantage. A notch below the US models.
One URL for every generation model I use (image, video, voice, avatar). One account, one wallet.
The latest release that crushes everything. Best on the market for photorealism.
State-of-the-art ultra-realism.
Quality on NB Pro's heels for 5× cheaper (ByteDance).
Unique stylistic consistency. Ideal for branding.
The artistic classic. Caught up on realism by NB Pro / Seedream.
Best at embedding text in an image (logos, posters).
Flux 2 ProBOpen source. LoRA possible. Very flexible.
N°1 voice right now. A prompt steers tone, accent, expression. No cloning.
Open source. Cloning that works wonders (Alibaba).
A leap forward. Bracketed directions ([laugh], [pause]). Huge voice library + cloning.
Voice bank, celebrities included. Check the rights.
200+ voices, 45 languages, Canva/Adobe integrations.
Cloning competitor, podcasters. A notch behind today.
The daily driver: 80% of my pro generations. Open source, 10-min one-shot, excellent lip sync.
Best quality I have tested. Period.
Open source, 4K, 50fps, audio sync in the same pass (Lightricks).
Unique multi-avatar feature. Animated 2-person AI podcast.
Kling just launched avatars. Behind on this vertical, overall Kling quality.
One of the first to make faces talk. ByteDance = TikTok = billions of training hours.
The classic. Clean, simple. Expensive, closed ecosystem.
A masterclass. 15 sec in 1080p. Unique: AI voice injection in the generation.
One of the best (Google). Expensive, and no voice injection (vs Kling).
Reference video + image = an animation with the same moves. Avatar that dances or presents a product.
Huge potential, TikTok base (ByteDance).
Open source (Alibaba), cheap, fine-tunable. Realism a notch behind Kling/Veo.
Aurora engine, 110,000 GPUs. Audio sync, decent photorealism (xAI).
Cheap, anime/stylized. Realism behind (MiniMax).
The absolute banger. Open source, self-hostable, total freedom over your workflows.
The B2B pioneer. Unavoidable on the enterprise side.
Excellent visually but losing the race (little code, server dependency).
Limited but built into M365. A huge advantage in the enterprise.
AI agent building.
Social media chatbot (Insta / Messenger / WhatsApp).
Voicebot / conversational chatbot.
Claude CodeSMy collaborator that executes. Not a chatbot: it acts directly on your files, folders, code.
NotebookLMAMy brain for documents. Your sources become an assistant that knows them. Largely free.
My meetings auto-transcribed. Invite the bot, no more note-taking. Very generous free tier.
AI voice dictation. Speech is 3-4× faster than typing.
AI focus music. Clinically proven: +119% beta waves.
My n°1 choice to host sites, apps, AI agents and my AIOS. Dead simple to start with.
Managed n8n solution.
The recipes
The prompts, techniques and formats shown during the conference. Everything is ready to copy.
# Action Write my reply to Martin's email. Goal: keep him. Offer a 15-minute call, with a refund only as a last resort. # Steps First acknowledge his frustration, then show him what he hasn't tapped into yet, then offer the call. # Persona You write in my place: Thomas Berton, founder of a Skool community about AI. # Examples My style, to imitate: "Hey Marc, good call on the feedback. You're right about module 1, too slow. Let me show you how to get straight to the point. Free Thursday?" # Context Martin has been a member for 3 weeks (490 euros). His email: the content is "too basic" and he wants a refund. The Tuesday live sessions answer his need exactly. # Constraints English, casual tone, short sentences. Zero corporate jargon, no superlatives. Under 150 words. No emoji. # Template Subject line + 3 short paragraphs + a closing question + signature: Thomas Berton, CEO Azuro.
Read and analyze the ASPECCT format below in depth. It is a tool for writing the best possible prompts for an AI model (LLM). When you have finished your analysis, reply only with "READY". Then, when I describe a need to you, you will write the best possible prompt for me following this format.
# ASPECCT summary
## ACTION
The action defines the mission by giving the AI an explicit task. This clarity of purpose produces targeted, useful results.
Example actions:
- Generate a market analysis report
- Write an engaging product description
- Develop a content plan for social media
- Create a list of article ideas for a tech site
- Write a sales pitch for a new piece of software
## STEPS
The steps give a sequence to follow. Structuring the process guides the AI toward the expected result systematically.
Example steps:
- Identify the campaign's target audience
- Study the competitors, their strengths and weaknesses
- Select the best channels
- Write punchy messages
- Propose visual ideas for each message
## PERSONA
The persona gives the AI a role. The chosen character filters the knowledge used and gives a voice and a perspective to the answers.
Example personas:
- Act as an experienced business consultant giving strategic advice
- Imagine you are a creative director in an advertising brainstorm
- Play the role of a financial analyst evaluating investment opportunities
- Give your advice like a motivational keynote speaker
## EXAMPLES
Show what you expect with concrete examples of the desired inputs or outputs. Examples serve as a reference to imitate.
Watch out: examples that are too precise can over-steer the model; sometimes vaguer or more numerous examples work better.
Example examples:
- Provide an executive summary of an old document as a basis for a new one
- Paste existing posts so the AI captures the voice and the tone
- Share a prospecting email that worked well and generate more like it
- Give your half-formed ideas: "I want a title that evokes an animal known for courage"
## CONTEXT
Give the environment, the circumstances and the relevant details. Context aligns the answers with the overall situation.
- Context of a product launch in a highly competitive market
- Context of a rebranding after a merger
- Context of a response to customer complaints on social media
- Context of a fundraising round with investors
## CONSTRAINTS
Constraints can be embedded throughout the prompt, or in their own section.
Example of Action + Constraint in the same sentence:
"ACTION: Write a short post under 280 characters."
The same prompt can also have a list of rules:
RULES:
- Output of 280 characters maximum
- Use only letters, numbers and common punctuation (. , ' " ?)
- Short, punchy sentences rather than long, wordy ones
Note: saying "don't do X" sometimes works badly (saying "no hashtags" contains the word "hashtags"). Prefer the positive phrasing ("use only letters and common punctuation"), or a reminder at the end of the prompt ("And remember: under 280 characters!").
## TEMPLATE
Define the form the output should take. An established format guides the structure and presentation of the generated content.
Example formats:
- Return your results in markdown
- Format the result in a plain text block
- Use this formula for titles: How to get {AWESOME} without {HASSLE}
- Label each result and justify your choice in bullet points
- Organize everything in markdown with headings, bullets and bold words
With ASPECCT in your toolbox, you can write prompts that produce creative, impactful results. When I give you a need, write the full prompt in ASPECCT format, in English, ready to copy and paste.
# Prepare my next meeting
## ACTION
Prepare my next external meeting from A to Z: identify the appointment and the guest in my calendar, research them, write a prep sheet, file it in the right place, link everything (calendar, reminder, email), and report back to me at the end.
## STEPS
1. Read my Google Calendar and find the next meeting with an external guest (a person whose email is not from my domain). Note the date, the time, the event title, and the guest's name and email.
2. Explore the guest's online presence: their website (the domain of their email is a good starting point), their LinkedIn page if you find it, recent news about their company. Keep what is useful for the conversation.
3. Write a prep sheet in a new Google Doc, with the following structure:
- Who: the person, their role, their company in 3 lines
- Context: why this meeting, any history of our exchanges
- Good to know: 3 to 5 key facts from your research
- Questions to ask: 3 to 5 relevant questions
- Desired outcome: what I need to walk away with
4. File this document in my Google Drive, in a "Meetings" folder (create it if it does not exist), named "Prep - {guest name} - {date}".
5. Add the document link to the description of the relevant Google Calendar event.
6. Create a Todoist task "Review the prep sheet - {guest name}" with a due time 15 minutes before the meeting starts.
7. Prepare a draft email in Gmail to the guest: confirmation of the appointment, a reminder of the time, one sentence on what we will discuss. Leave it as a draft.
8. Finish with a report: what you found, what you created, and the links to the doc, the event and the draft.
## PERSONA
Act as an experienced executive assistant: rigorous, concise, who anticipates needs and leaves nothing to chance, but who never makes a committing decision in my place.
## CONTEXT
You are connected to my tools: Google Calendar, Google Drive, Gmail and Todoist. You can read my calendar and my documents, create files and tasks, and prepare emails. The target meeting is the next external meeting coming up, not an internal team meeting.
## CONSTRAINTS
- All your output is in English.
- The email stays a draft: you never click send, I send it myself.
- You never delete or move any existing file or event.
- Before any irreversible or ambiguous action (two candidate meetings, guest not found), ask me instead of guessing.
- Base the sheet on real sources; if a piece of information is uncertain, flag it rather than inventing it.
{
"subject": "cinematic portrait",
"action": "standing confidently",
"context": "next to a vintage car, golden hour",
"style": ["cinematic", "moody"],
"camera": { "lens": "85mm portrait lens", "depth_of_field": "f/1.4" },
"lighting": ["golden hour lighting", "rim lighting"],
"film_look": ["Kodak Portra 400"],
"quality": ["ultra-detailed"]
}
Professional character reference sheet, technical orthographic views, single sheet, clean layout. LEFT SIDE: full body reference, neutral anatomical pose, arms slightly away from the body, matte charcoal athletic suit, white seamless background. RIGHT SIDE: two portraits, stacked. TOP: front headshot, neutral expression, exact face from the input image. BOTTOM: 90 degree left side profile, matching lighting and scale. Consistent studio lighting, no shadows on the background, no text, no watermark, 4K, photorealistic quality.
# Mac / Linux curl -fsSL https://claude.ai/install.sh | bash # Windows (PowerShell) irm https://claude.ai/install.ps1 | iex # Then, from any folder: claude
Stress-test your offer before a real client.
It gathers context instead of guessing.
A harsh, honest critique of your work.
Paste it into Claude to watch the reasoning unfold.
No need to write a structured prompt by hand: have a meta prompt write it. A real ASPECCT prompt is worth it when the task is reused, high-stakes, and must stay consistent.
"Don't do X" works badly (saying "no hashtags" contains "hashtags"). Say what you WANT, or add a reminder at the end of the prompt.
Examples that are too precise over-steer the model. Sometimes vaguer or more numerous examples work better.
Against hallucinations, connect the LLM to the web (Perplexity) or to your documents (NotebookLM, Projects).
All the models in one place, filtered by verified sources (health, finance, science), under 20 euros a month.
Same engine as Cowork, but the consumer wrapper caps the power. Better to learn where everything happens.
Ask it to sort out a messy folder (e.g. Downloads). Simple, telling, immediate.
Free conversion sites sell your data. Claude Code + ffmpeg = all local, nothing uploaded, free, no limit.
Give the link to Claude Code: it downloads, transcribes and returns the structured takeaway. You do something else.
Launch the mission in the evening (download, transcription, sub-agents, vector base), check in the morning. 60h of video processed, ~1h of human time.
Nano Banana Pro (image) + ElevenLabs (cloned voice, 15s) + HeyGen (avatar), then Veo 3 for the scene and the sound.
A single input photo, several settings, the same face intact.
With vision, show a photo (fridge, error screenshot, whiteboard diagram) instead of writing it all out.
Today, offering an MCP means being pluggable, so competitive.
Lens and depth of field: "lens": "85mm portrait lens", "depth_of_field": "f/1.4", angle "close-up" or "medium shot".
"film_look": ["Kodak Portra 400"], "soft film grain", "quality": ["ultra-detailed", "sharp focus", "8k"].
Single sheet, orthographic views, front portrait + 90-degree profile, white seamless background, 4K, no text, no watermark.
Avatar renders shown in an iPhone mockup (~9:18.5), photo cropped like a camera.
The video players are in aspect-video (16:9), the standard for YouTube and the big screen.
ElevenLabs clones a voice from a 15s sample. HeyGen makes the video clone from 15s filmed on your phone.
Want to go further?
Revisit the presentation, try AIOS, or join the community.