From Prompt Tricks to Real AI Workflows

Michael Lee, CRO | Data & AI | Scaling $1M-$100M B2B Companies With AI | Turning Lean Teams Into High-Output Engines with Agents + Systems | Top 2% worldwide, recently posted something that made me stop scrolling: "Prompt engineering is dead. And 99% of people are still learning yesterday's skill. Two years ago, everyone became a prompt wizard."

As Michael Lee pointed out, we spent the last couple of years obsessing over magic phrases: "Act as a Senior VP...", "Think step by step...", long prompt templates that felt like cheat codes. We thought, as he put it, that these magic phrases "unlocked intelligence." But we were wrong.

"Prompt engineering is dead."

"And 99% of people are still learning yesterday's skill."

"Two years ago, everyone became a prompt wizard..."

When Michael Lee posted this, he was not saying prompts no longer matter. He was saying something much more important: the real leverage has moved from single prompts to flows—structured, agentic workflows that turn raw model intelligence into consistent, scalable outcomes.

The Data That Changes the Conversation

Lee references Andrew Ng and a simple but powerful benchmark story. On HumanEval, a standard coding test, here is what happened:

ChatGPT previous gen (zero-shot): 48.1%
ChatGPT new gen (zero-shot): 67.0%
Previous gen + an agentic workflow: 95.1%

In other words, an older model with the right workflow beat a newer, smarter model used in a naive way.

That line from his post hits hard: "An old model with a workflow outperformed the new GPT model. Read that again." The implication is huge: model choice matters, but workflow architecture matters more once you are past a basic quality bar.

If you are still trying to squeeze results out of one giant prompt, you are effectively stuck in 2023.

From Prompt Engineering to Flow Engineering

Lee frames the shift like this:

2023: prompt engineering
2025: "Why isn't this scaling?"
2026: flow engineering

I think he is exactly right about that middle part. This is where a lot of teams are today. They have clever prompts, proof-of-concept demos, maybe even some internal tools. But they quietly ask themselves: Why isn't this scaling across the business? Why is everything still fragile, manual, and ad hoc?

Flow engineering is the missing layer.

Where prompt engineering focuses on what you say once to the model, flow engineering focuses on how work moves through a sequence of steps, checks, tools, and agents to reach a reliable outcome.

Prompt Engineer vs Flow Engineer

Michael Lee uses a simple example:

The Prompt Engineer says:
"Give me a go-to-market strategy for my SaaS."
Output: generic MBA fluff.

The Flow Engineer designs something very different:

Analyze competitor pricing
Extract customer pain points
Map feature gaps
Draft positioning angles
Score each against data

Output: a strategy you can actually execute.

Same underlying model. Completely different architecture.

The flow engineer is not trying to write a magic prompt. They are decomposing a business problem into steps, assigning each step to an AI agent or tool, and then orchestrating how those steps connect. This is exactly how real teams work; we are just starting to do it with digital teams.

We Are Not Chatting with AI Anymore

One of the most important lines in Lee's post is this:

"We aren’t chatting with AI anymore. We’re managing digital teams. We’re chaining agents. We’re designing oversight."

This is the uncomfortable truth for anyone still obsessing over clever prompt templates. The skills that matter now look a lot more like management, operations, and systems design than like copywriting cleverness.

If you lead a sales, marketing, or operations team, that should actually feel familiar. You already:

Break goals into projects and tasks
Assign work to people with different strengths
Put in QA and approval steps
Track performance and iterate the process

Agentic AI is just the digital version of this. Instead of a human SDR or analyst, you might have:

A research agent that pulls competitor and customer data
An analysis agent that clusters pain points and segments
A strategy agent that drafts options
A QA agent that checks for gaps or hallucinations
A reporting agent that packages everything for your CRM or slide deck

Your job is no longer to "chat" with a model. Your job is to manage this digital org chart.

What Flow Engineering Looks Like in Practice

So what does it mean to embrace flow engineering in day-to-day work? Building on Lee's argument, it looks like:

Start from the business outcome, not the prompt.
Instead of asking "What prompt do I use?", ask "What recurring outcome do I need every week or every day?" Sales sequences, product research, client reports, pipeline analysis—these are flows.
Decompose the outcome into repeatable steps.
Each step should be simple enough that an agent can do it with clear inputs and outputs. For example: "Given this call transcript, extract top 5 pain points and tag them with product features."
Assign each step to an agent or tool.
Sometimes it is a single LLM call. Sometimes it is a mix of LLM + API + database. The point is that you stop trying to make one overstuffed prompt do everything.
Add oversight and feedback loops.
Flow engineering is not "set and forget." You design checkpoints: automatic validation, human-in-the-loop review, or secondary agents that critique the output.
Measure, then refine the architecture.
Just as you would improve a team process, you adjust steps, tools, and agents based on quality and speed. You are engineering a system, not polishing a sentence.

This is where Lee's framing of "Agentic AI is the new org chart" becomes practical. The winners will not be those with the prettiest prompt libraries; they will be those who design AI-first workflows that plug directly into revenue, cost savings, and customer experience.

The New AI Management Skill Set

If prompt engineering is yesterday's skill, what replaces it?

Drawing from Michael Lee's post and what I see in the field, the emerging skill set looks like this:

Problem decomposition: Turning a fuzzy objective into a clear sequence of steps.
Workflow design: Knowing when to use humans, when to use agents, and how they interact.
Data awareness: Understanding what data each step needs and where it lives.
Oversight design: Building in guardrails, escalation, and QA.
Change management: Helping real teams adopt these flows without breaking everything.

You do not need to be a deep engineer to do this. But you do need to think like a manager of digital talent, not just a clever prompt writer.

Stop Learning How to Prompt, Start Learning How to Manage

Michael Lee ends with a clear call to action: stop learning how to prompt, start learning how to manage. I think that is the right north star for anyone serious about AI in 2025 and beyond.

Yes, you still need to know how to write a clear prompt. But that is table stakes now, like knowing how to write a decent email. The leverage has moved up a level—to how you architect flows, chain agents, and wrap all of it in a system that your business can trust.

If you feel like your AI efforts are not scaling, do not look for the next magic prompt. Take another look at your workflow.

Ask yourself Lee's implicit question: am I chatting with a toy, or am I managing a digital team?

The answer will determine whether AI is a party trick in your organization—or a real competitive advantage.

This blog post expands on a viral LinkedIn post by Michael Lee, CRO | Data & AI | Scaling $1M-$100M B2B Companies With AI | Turning Lean Teams Into High-Output Engines with Agents + Systems | Top 2% worldwide. View the original LinkedIn post →