Dylans mange gode råd
Dylan er en gut, der nu og da sender gode råd ud i et meget kort nyhedsbrev. Hans firma er her: Gradient LabsJeg har samlet dem her. De er på engelsk, men I kender jo nok en AI, der kan oversætte det til dansk, hvis I ønsker det :-).
Identity.
Your AI performs 10x better when you give it a specific identity. Think of it like casting an actor for a role - the more specific the character, the better the performance.
Here's the framework:
Step 1: Pick a persona (doctor, copywriter, Seneca)
Step 2: Place it at the TOP of your prompt
Step 3: Give that persona a clear objective
Examples:
Writing: "You're an expert communicator who favors simplicity. Rewrite this email for clarity."
Relationships: "You're a therapist trained in Esther Perel's methods. Answer this relationship question."
Taxes: "You're a business tax expert for S-corps in Texas. Find tax-saving opportunities."
Pictures. Smart creators use a two-model strategy.
For original images: GPT-4o Imagen beats everything else at creating a variety of images from scratch.
Pro hack: Ask ChatGPT to research prompting best practices for GPT-4o Imagen, then have it create your prompt. Your images will be 10x better.
For charts and dashboards: Claude Sonnet/Opus builds beautiful interfaces with code that you can screenshot for presentations.
The combo: Use GPT-4o for creative assets, Claude for data visualization and UI mockups.
Claude's new deep research feature is now the king (or queen) of deep research. While OpenAI reads 80-100 sources, Claude tears through 250-500 sources per search.
The difference? Quality and depth.
Claude doesn't just skim - it synthesizes multiple perspectives across way more sources and writes better summaries. This is because in the background they’re using a multi-agent, not single-agent architecture.
Quick test: Next research task, try Claude's deep research instead of your usual tool. The extra sources create insights you'd miss otherwise.
Ask 3 times. Most people ask AI to do something once, get mediocre results, then complain AI sucks. Here's what actually works: The 3x Rule.
Ask the same question 3 different times. Get 3 different outputs. Pick the best parts from each.
Example: Need a sales email? Don't ask once. Open 3 tabs, ask with the same prompt 3 times, compare results. Then feed all 3 into another AI asking it to combine the best elements.
Why this works: AI is probabilistic (fancy word for unpredictable). Same input = different outputs every time.
Pro tip: For quick broad questions, stick with OpenAI (without deep research turned on). For thorough analysis, Claude wins.
Landing pages.
Most people open Claude and type "build me a landing page." The AI spits out something that works. But it looks like every other AI-generated site on the internet. Generic. Forgettable. Low-converting.
Here's the problem: You skipped the research step.
Before asking AI to build anything, have it research what actually makes landing pages convert as of today's date. Copywriting tactics. Page structure. Color psychology. The works.
Then—and this is the key—feed that research back to the AI in the same conversation when you ask it to write your copy and build your page.
Why does this work? Because now the AI has context. It's not defaulting to the most common template in its training data. It's applying current best practices to your specific product.
One extra prompt. That's the difference between "meh" and "this actually converts."
I made a video walking through my full process if you want to see it in action—research, copy, building, iterating, deploying. No code required.
Meta-prompting. The Simple Method: When you need a better prompt, ask AI to write it for you.
Example: Instead of struggling with a prompt, send this: "Can you write a prompt for me to help [your task] following current best practices as of [Today’s month/year]? Research those practices first, then improve my original prompt."
Why this works: AI knows the latest prompting techniques better than you do. It's literally designed for this.
Bonus: Use Anthropic or OpenAI's prompt generators. Drop in your basic prompt, watch it transform into something 10x better.
Document first, automate second.
The Process:
What do you want from your business? (More clients? Better throughput? Less churn?)
What activities revolve around that primary goal?
Which of these is your current bottleneck?
Break that bottleneck into individual functions, then map each step like someone strapped a GoPro to your chest
Example: Don't just say "automate client onboarding." First document every step: "Send welcome email → Schedule kickoff call → Collect intake form → Create project folder."
AI image prompts.
Here’s what actually works: Give AI five inputs and let it write the prompt for you.
The five inputs:
Purpose (Instagram ad? YouTube thumbnail? Landing page?)
Audience (who’s seeing this?)
Subject (product, person, scene)
Brand rules (colors, style, vibe)
Reference image (optional)
That's it. You act as creative director. AI handles the technical translation into prompts optimized for Nano Banana Pro.
I built a system prompt that takes these five things and spits out three ready-to-use prompts. You just copy, paste into Nano Banana Pro, generate.
AI writes in YOUR specific style, tone, and structure?
Phase 1 (Build Your Style Database): Upload 10+ writing examples to a Claude/GPT project. This becomes your reference library.
Phase 2 (Create Your Style Analyzer): Ask AI: "What elements do experts analyze when studying writing style?" → You'll get a list (sentence length, word choice, rhythm, etc.) → Feed this list to another AI and ask: "Create a prompt that extracts these elements from writing samples"
Phase 3 (Analyze Your Examples): Run your uploaded examples through that extraction prompt. → You'll get a detailed breakdown of the writing patterns
Phase 4 (Build Your Style Cloner): Take that pattern breakdown to a final AI and say: "Create a rewriter prompt that mimics these exact patterns."
Phase 5 (Clone Away): Use your custom rewriter prompt on any draft. Boom - instant style matching.
Clone perfectly,
Treat AI like a board, not a search engine
The top 1% treat it like a custom board of advisors.
Here's what they know that you don't:
Projects. Not prompts.
Every major AI (ChatGPT, Claude, Gemini) lets you create specialized AI assistants loaded with your context, files, and custom instructions.
Think of it like hiring a tax expert who already knows your P&L, incorporation details, and current strategies. Or a proposal writer who understands your services, tone, and client history.
The magic happens when you:
Upload relevant files (your data, preferences, examples)
Write custom instructions (expertise, persona, perspective)
Use the best available model (not locked into outdated versions)
I've built AIs for taxes (saved $10k), proposals (hours to minutes), and even relationship advice (bottled Esther Perel's wisdom).
Ready to build your AI advisory board?
Start with one task you do weekly. Create a project. Upload context. Watch the magic.
Build your board,
Finding the "best" AI model.
Wrong question.
The real question: How do you extract maximum value from ALL of them?
Here's the model arbitrage method:
Step 1: Ask your research question to 4 different providers:
Grok (for real-time X insights)
Perplexity (finds sources others miss)
ChatGPT (comprehensive coverage)
Claude (best overall quality)
Step 2: Each gives you different angles, sources, and covers blind spots.
Step 3: Feed all responses into one quality reasoning model (Claude Sonnet/Opus or GPT-5) with this prompt: "Synthesize these 4 research outputs into one definitive report."
Result? You get 4x the perspectives, sources, and insights.
Most people pick one model and call it done. Smart operators play model arbitrage.
Talk to the AI
Here's the simple trick: Start dictating to AI instead of typing.
I use dictation for 50% of my AI interactions now - coding, research deep-dives, data analysis. The speed difference matters.
Three easy ways to start:
Native: Windows and Mac have built-in dictation
Mobile: ChatGPT and Claude mobile apps have excellent dictation
Desktop: Tools like Superwhisper and Wisper Flow work across any platform
Why this works: You can rant for 2 minutes about a complex problem instead of typing 5 paragraphs. AI gets better context, you save time, everyone wins.
Pro tip: Don't worry about dictation mistakes like "Claude" becoming "cloud" - AI figures it out from context. Unless you’re coding, fix it when you see it. You can thank me later.
Content creation
Most people treat content creation like throwing spaghetti at the wall. They post random thoughts and wonder why nobody cares.
Here's the system I use to turn podcasts and YouTube videos into high-performing social content:
The 4-Step Content System:
Collect Raw Material - Save links to podcasts/videos that make you think "this is useful." Add 3-5 words describing why. Aim for 8-10 pieces weekly.
Extract the Transcript - Paste the YouTube link into Gemini with: "Pull the transcript from this video and give me the text." Takes 30 seconds per video. If that doesn’t work, have GPT/Claude write a python script using yt-dlp.
Build Your Writing Assistant - Find one person who writes great posts on your platform of choice. Feed their style guide or best posts to ChatGPT/Claude and ask: "Create a system prompt that writes like this person."
Feed and Polish - Drop your transcript into your GPT/Claude Project with the previously created prompt: "Turn this into a post about [key insight]." Then tweak the output by adding rules in the system prompt like "no buzzwords", "end with questions", “write at a 5th grading reading level.”
You're not just reposting—you're adding your taste and perspective to valuable content that already exists.
This works for LinkedIn, X threads, newsletters, even Instagram captions. The format changes, but the process stays the same.
Doing taxes with AI
Your accountant is smart. But they're also human.
They forget things. Get busy. Miss new strategies.
AI doesn't.
Here's how I built a tax-saving AI that found $10K my CPA missed:
Step 1: Upload everything to Claude Sonnet/Opus - last 3 years of returns, bank statements, business docs, investment accounts. Give it your full financial picture.
Step 2: Ask: "Research current tax strategies for someone in my exact situation as of [today's date]." This forces it to pull the latest tax code changes.
Step 3: Take the AI's suggestions to your accountant. Don't implement alone - confirm first.
Kommentarer
Send en kommentar