Workflows

Everything so far has been one conversation at a time. Today is the step where a good conversation becomes a system that produces the same quality without you being clever again each time. This is a real field with named patterns that people build on — so I'll teach you those first, then show you mine as one worked example of one of them.

1. One good chat doesn't compound. A workflow does.

Think about what you did on Day 4. You wrote five checklist items, ran a review, got a better piece of writing.

Now answer honestly: could you get the exact same result next Tuesday? Same standard, same catches, without remembering what you typed?

Most of you couldn't — not because you did it badly, but because it lived in a chat window that is now gone.

A workflow is the same steps, the same context and the same quality bar, written down somewhere you can run again. That's the entire idea. Everything else today is about how to structure one.

2. Chat, workflow, automation, agent — and how they actually relate.

These four words get used interchangeably, including by people selling you things. But they are not four options on one line, and treating them that way is the first thing that confuses people.

There are two separate questions underneath.

Question 1 — who decides the steps?

Workflow	You do, in advance. The path is known before you start.
Agent	It does, as it goes, based on what it finds.

Question 2 — who runs them?

Manual	You, by hand. You start each run.
Automated	A trigger. A schedule, an email arriving, a form being filled.

So automation is not a third kind of thing. It is how a workflow runs, not an alternative to one. The same workflow run by hand on Monday and on a schedule on Tuesday is the same workflow — one of them is automated.

Which means the two questions combine:

	Run by hand	Automated
Workflow	You follow your own written steps.	Those steps fire on a trigger without you.
Agent	You give it a goal and watch it work.	It receives goals and runs on its own.

Chat sits outside both questions. Nothing was defined in advance and nothing repeats — which is exactly why it does not compound.

The useful rule for question 1: use a workflow when you know the path in advance. Use an agent when the next step genuinely depends on what it finds.

Most tasks you will meet in marketing have a known path. That means most of the time the right answer is a workflow, not an agent — which is the opposite of what the current hype will tell you.

And for question 2: automate only once the manual version works. Agents are tomorrow.

3. The five patterns almost every workflow is built from.

This is the core of today. These five are the common vocabulary — the naming here follows Anthropic's Building Effective Agents, which is the reference most teams have standardised on, and you'll meet the same five under the same names elsewhere.

They are composable. Real systems nest them inside each other.

One word first, because it appears in all five. A "call" is one round trip to the AI — you send something, it sends something back. Typing a message into ChatGPT and reading the reply is one call. A program doing the same thing is an API call; same idea, different sender.

Each call is separate, and from Day 1 you know the model remembers nothing between them. That is exactly why these patterns are built from several calls rather than one long conversation — a fresh call is a fresh reader, which is sometimes the point.

1. Prompt chaining

Break a task into steps. Each step's output becomes the next step's input.

Use it when a task has a natural order and each stage is easier to get right in isolation. The check between two steps is called a gate, and it's the cheapest quality control that exists — catching a bad outline costs seconds, catching a bad article costs an afternoon.

2. Routing

Classify the input first, then send it to a different specialised prompt depending on what it is.

Use it when one prompt is being asked to handle cases that need genuinely different treatment. A single prompt trying to serve all three ends up mediocre at all three.

3. Parallelisation

Run several calls at the same time. Two flavours:

Sectioning is for speed and for stopping one output bleeding its style into another. Voting is for catching mistakes — and it is a version of what you already did on Day 1 with five assistants.

4. Orchestrator–workers

One call breaks the job into subtasks and decides what they are, hands each to a worker, then assembles the results.

The difference from sectioning: here the number and shape of the subtasks isn't known in advance. That makes this the pattern closest to an agent, and the first one where you lose some predictability.

5. Evaluator–optimizer

One call produces the work. A separate call criticises it against a standard. The work goes back for another pass. It loops until it passes.

Use it when you have clear criteria and when a critic can reliably say what's wrong. This is Day 4's "AI reviews first" turned into a loop, and it's the pattern my own system uses — section 6.

4. The anatomy: context block, acceptance criteria, approval gate.

Whichever pattern you pick, three parts decide whether it's actually usable.

The context block

You've all felt this: every new chat, you re-explain the company, the customer, the tone. By the fourth time it's tedious, so you cut corners, and the output gets worse for reasons that have nothing to do with the AI.

Fix it once. Keep a file with what never changes — what the company does, who the customer is and what they fear, three examples of writing you'd publish, banned words and claims, what the piece must achieve.

Paste it at the top of every run. It's Day 2 made permanent instead of retyped. The test of a good one: a colleague could paste it in and get output as good as yours.

Acceptance criteria

How will you know it's actually done? Not "write the post" — what has to be true about the post before anyone calls it finished.

Write these before you start. They're what an evaluator step checks against, and you cannot ask for a self-review if nothing was ever written down to review against.

The approval gate

A point where nothing proceeds until a human says yes. Not a glance — a decision by someone who could say what they checked.

You need one before anything that reaches a customer, spends money, or carries your name. This is the step that does not move no matter how good the rest gets.

5. Get the plan right first. A well-run wrong plan is still wrong.

This is the part I personally spend the most time on, and the part almost everyone rushes.

When I build something, I don't start by asking for the work. I ask for a plan, and I give far more input than feels necessary — what I'm building, what I actually need, what the finished thing must do, the ultimate goal rather than just the immediate task, and the constraints.

Then I read the plan properly before any real work begins.

The reason is simple arithmetic. Perfect execution of a wrong plan is still a wrong result, and you've now spent hours getting there. Every bad output I've had from these tools traces back to a plan I approved too quickly.

I also use the strongest model available for this step specifically. Planning is where the difference between models actually shows in the result; for the work afterwards it matters much less.

6. A worked example: my review system is an evaluator–optimizer.

On Friday I told you lots.blog is back and every post goes through an AI review before a human reads it. I said twenty checklist items. It's nineteen — and now you have the vocabulary to see what it actually is.

It's pattern five. A generator writes, an evaluator scores against nineteen weighted criteria, and it loops until it passes. Nothing exotic — a standard pattern, applied to a specific problem.

Weight	What it checks
High	AI slop / originality · Title promise · Headline quality · Reader outcome · Opening hook and context · Strategy alignment · Search intent · AEO answer quality · Current evidence · External sources · Specific detail
Medium	Structure · SEO on-page · Internal links · FAQ opportunity · Structured data · Readability
Low	Next step · Publish polish

The weighting is the design decision worth copying, not the specific items. Originality, whether the title keeps its promise, whether the reader can do something afterwards, and whether claims are sourced sit at the top. Formatting and SEO sit below them.

That ordering is a direct correction of what killed the first version of the blog. Those old automated posts would have scored well on structure, headings and readability. They failed on exactly the things I now weight highest.

Your criteria will not look like mine, because you're not publishing what I publish. What transfers is the shape: weighted criteria, checked by a separate step, looping until it passes.

7. One real review, in full.

Here's actual output from that system, unedited.

It passed seventeen. What matters is the two it failed — both high-weight, and both the same underlying problem.

Three confident-sounding statistics with nothing verifiable behind them. A citation naming a real organisation but linking to nothing. A reference silently truncated mid-sentence.

That's Day 1, turning up in my own work six months later — in an article that scored 88% and that most readers would have read without noticing.

Two things to take from this. An evaluator only catches what its criteria describe: those failures were found because "current evidence" and "external sources" are items on the list. And notice the passing entries quote the specific thing that earned the pass — "the fitness studio vs financial advisor content themes" — because a review that can't point at evidence isn't a review.

8. Where the tools are — and when you don't need one.

You can run every pattern above in browser tabs with copy-paste. For learning, you should — it's how you find out where a workflow actually breaks.

When you want it running on a trigger rather than by hand, that's the automation layer, and there's a crowded market for it. Worth knowing the names so you can evaluate them yourself: Zapier, Make and n8n come from general automation and have added AI steps. Gumloop, Lindy and AirOps were built around AI workflows from the start. Several have free tiers, and n8n can be self-hosted.

That list will go out of date, and it already has. An earlier version of this page recommended Relay.app. It shut down in 2026 — free accounts on 15 August, paying customers on 14 September — after its founder was hired by Google. Nouverah caught it and told me.

Which is the actual lesson here, and it is worth more than the list: this market churns. Tools close, get acquired, or change their pricing every few months. That is another reason to write the workflow down somewhere neutral before you commit it to any platform — the workflow is yours, the platform is rented.

I'm deliberately not recommending one. They change constantly, your needs aren't mine, and picking a tool before you've written the workflow down by hand is how people end up automating a bad process.

9. What this lesson did not cover, and how to go deeper.

I want to be straight with you about the size of what you've just read. This is an introduction to a field, not the field.

What we covered: the four terms, the five composable patterns, the context block, acceptance criteria, approval gates, and one worked example.

What we did not, and you'll meet as soon as you build anything real:

Where to go next. Anthropic's Building Effective Agents is the source of the five patterns and worth reading in full. The platforms named above have their own documentation with working examples. And the honest answer is that this field is a few years old and moving fast — treat any single source as one view, including this lesson and including me.

What I can promise you is the shape. The vocabulary here is stable, the patterns are what people actually build with, and if you can name which pattern a system is using, you can read almost any tutorial you find and place it immediately.

Tomorrow — Day 6: Agents. What happens when you stop defining steps and hand over a goal instead — how agents actually work, when they're the right answer, and why most agent projects fail.