Most people use AI the same way they use a search engine
You type something in, you get something back. Useful. But limited.
An agentic workflow is different. Instead of answering a single question, an agentic system works through a task, gathering information, making decisions, using tools, producing real output, with minimal hand-holding.
You ask, it answers, iterate, start over.
Gathers, decides, uses tools, produces output.
Most people sit firmly in the traditional AI space. Meanwhile, the industry and its state-of-the-art models are pushing hard towards the agentic side. There is a significant gap between the two, and bridging it is exactly what this series is for.
Every AI workflow has four building blocks
Whether you're drafting a pitch deck, automating a report, or building a presentation from scratch, every workflow is made of the same four components. The platforms change. The principles don't.
LLM & Environment
The intelligence and the workspace where you interact with it.
Memory & Knowledge
What the AI knows beyond its training, your data, your context.
Tools
Extra capabilities that let the AI act in the world, send, fetch, trigger.
Instructions
How the AI thinks, what it produces, and the rules it follows.
You've already used all four
You've used an LLM. You've added a PDF to give it context. Some of you have connected it to Gmail or Drive. And you've definitely written a prompt.
What this series does is take those instincts and turn them into a system, deliberately, level by level.
Each building block has three levels of depth. The next section shows you the progression, and where this series will take you.
Three proficiency levels
Where are you now, and where can you go?
Using standard interfaces and defaults. Browser-based chat, manual input, one-off prompts. Useful, but limited by the ceiling of each session.
Specialised environments, deliberate configuration, and structured workflows. Moving from default settings to purposeful, repeatable systems that produce consistent results.
Developer-grade depth, multi-system orchestration, and full customisation. Building complex, automated workflows that coordinate across multiple AI systems and external services.
The four blocks, level by level
The LLM, the Large Language Model, is the intelligence at the centre of everything. ChatGPT, Claude, Gemini: these are different LLMs. Different training, different strengths, different character.
The environment is where you work with that LLM. Same model, very different experience depending on the interface. Claude in a browser chat window is not the same thing as Claude running inside VS Code. The LLM is the engine. The environment is the cockpit.
ChatGPT.com, Claude.ai, Gemini, NotebookLM. You open a browser, type a message, get a response.
Claude Code in your terminal or inside VS Code, Claude Cowork, Antigravity. Purpose-built applications like Lovable, Framer, Make, Base44. The AI can see your files, work across your projects, and take actions, not just produce text.
Claude Code used to its full technical depth, Codeword, n8n for connecting multiple LLMs and building complex automated workflows.
LLMs don't know anything about you. They were trained on enormous amounts of public data, but your documents, your clients, your projects, none of that is in there. Memory is how you close that gap.
Think of it as briefing a new colleague before they start. Without a proper briefing, even a brilliant person produces generic work. With one, they can get straight to the point.
Paste text into the chat, upload a PDF, attach a document. Context is provided manually, one conversation at a time.
In Claude Code, working inside a project directory means the AI reads your file tree automatically. No re-uploading, no re-briefing each session. Your context persists across every interaction.
Cloud tools like Supabase for structured data, vector databases like Pinecone that let the AI search your documents with precision. A proper indexed library rather than a folder of loose files.
Think of tools as giving AI arms. Each one extends what the AI can do beyond generating text. Without tools, the AI is contained within its own text window. It can write, but it can't act.
The underlying mechanism that connects most of them is something called MCP, the Model Context Protocol. Think of it as a universal adapter: a standardised way for AI to plug into external apps and services.
Gmail, Outlook, Google Drive, SharePoint, basic web browsing. The AI helps you use these services, but you are doing the connecting.
Direct, live connections between the AI and your services. Instead of helping you draft a Slack message, it sends one. This is where the AI shifts from assistant to agent.
GitHub, Vercel, custom API integrations. Platforms like n8n, where multiple AI agents hand work between each other, each handling a different part of a workflow.
Instructions govern how the AI behaves, what role it plays, how it communicates, what it does when it's uncertain, and how it produces particular outputs. Think of them as the control layer, shaping how the LLM uses its memory, its tools, and its own reasoning.
You can have the best model, a rich knowledge base, a full set of tools, and still produce inconsistent output if your instructions are vague.
Give the AI a role, context, objective, and format. "You are a senior copywriter. Here's the brief. Write a 300-word product description in a professional tone."
A permanent briefing carried across sessions. In Claude Code, this is a CLAUDE.md file defining project conventions, feedback loops, self-checking, and how the AI performs certain actions.
Custom, self-contained instruction packages. The AI recognises the task, activates the right skill, and delivers consistent, high-quality work with domain-specific precision.
Where are you now?
Around 95% of people are operating at Level 1 across most or all of these components. Level 1 is useful.
If you're reading this, you already sense there's more.
Move you to Level 2, practically and consistently, across all four building blocks. We will touch on components of Level 3 as we go along the examples.
The framework transfers
Every piece in the series takes the four components one level deeper. Real workflows, real examples, something you can act on immediately.
What you learn with Claude works in ChatGPT. What you build in one environment carries to another.
Lowest tier subscription, £18/month.
Free to download. Where we build.
No complex setup, no technical background required. The constraint is deliberate.