Swift ✉️ Memos

Is there a better way than prompting to get what you want from AI?

The most advanced AI users have moved to creating loops. But what about the lighter, everyday use cases that don’t require such a heavy-handed approach?

There are simpler tactics that provide an immediate upgrade to the endless “change this” and “change that” prompting style that every AI user starts with.

The tactic I want to talk about today can be applied to anything you turn to AI for: text, images, code… anything you’re expecting an output on.

Instead of refining a single output five times over until it reaches the point of “good enough”, I’d like to encourage you to try the fan out method.

So today, I want to break down:

How our work habits are outdated relative to the cost of production
What the fan out method is (and why it’s applicable to everything you do with AI)
How to implement the fan out method (prompt included)

—@dharmesh

The New Cost of Production

It wasn’t long ago that producing usable work of any sort was a long, expensive process. In that BC era (Before ChatGPT), the time it took to create a rough draft was simply the time it took. There weren’t many shortcuts available.

Building a working internal tool required an engineer. Writing placeholder landing page copy required a junior copywriter. The cost of production was relatively high, which affected how workflows were designed.

There are invisible scripts about what efficient work looks like that we follow today only because we haven’t yet fully adapted to the power of AI. One of these scripts is “production is expensive” -- and that’s the script we’re going to explore today.

Put yourself in the shoes of an old-school film photographer. Film is expensive and time-consuming to develop. Before the shutter is clicked to take a photo, they have to have a high degree of confidence that the frame is exactly how they want it to be, else risk a wasted photo.

Now contrast that with what it’s like to take photos on an iPhone. Photos cost nothing but a few megabytes on phones that now hold terabytes. Professional photographers can now fire off a thousand snaps at a wedding to maximize their chances of capturing the perfect moment.

They’ll only keep, say, the best 50 photos. But because it costs them virtually nothing to capture an additional frame, they shoot for volume.

With AI, it costs virtually nothing to create first drafts. So how can we best take advantage of that in a practical way?

What Is the Fan Out Method?

We can best take advantage of the low cost of generation by generating in bulk, then keeping the best option as the base draft. That’s the fan out method.

The term comes from the world of electronics originally, but was later adopted in computing to represent a “one-to-many” distribution. A fan out is essentially when a single input is spread across many parallel paths at once.

In practice this means asking AI to generate multiple versions of an artifact for you to pick from instead of just one. Instead of asking for a single paragraph for the article you’re writing, you ask for five variations.

There are plenty of good reasons to do this.

Refining over and over has its limitations. I previously wrote about how asking the same AI that produced the work to also edit the work is a flawed approach.
Comparing is easier than critiquing. It’s easier (and faster) to pick between multiple options than spend time auditing a single draft.
Variety helps identify the direction you actually want. The fan out method can produce options you wouldn’t have even thought to ask for in the first place.

Another meaningful benefit: picking one option out of many reveals preferences that you may not have been capable of putting into words.

Graphic designers famously express frustration about impractical feedback like “make it pop more”. While I understand their frustration, it is genuinely difficult to give direction in a domain where you lack the vocabulary and expertise.

Mark this down as another win for the fan out method. Making ten versions of the same design is a lot of added work for a human designer. But generating ten versions (instead of one version) is trivial work for a modern AI model.

Trying this yourself is very simple. Here’s a quick example prompt:

Create 5 different versions of a LinkedIn post summarizing the most recent simple.ai newsletter. Make 2 of them long posts, 2 short posts, and 1 post optimized to do well on LinkedIn based on best practices. Approach it from different angles and try different styles to get some variety.

Number the versions you generate so I can easily ask you to iterate on one or combine elements of one with another.

The goal here isn’t to one-shot a perfect ready-to-post output. It’s to find a stronger first draft in less time than it would take to polish a single output.

Try It Today

The fan out method is a direct response to the costs of generation reaching new lows. And, as I covered above, it doesn’t need to be perfect to be useful.

What’s nice about the fan out method is how simple it is to implement. This is a tactic beginners can immediately understand and put to use.

Looking ahead, it seems reasonable to predict that generating artifacts is only going to get cheaper and faster from here. As more people generate more things, taste and judgment will grow increasingly valuable.

The next time you ask AI to create anything for you, try being a little greedy and ask for multiple versions of the same thing.

—Dharmesh (@dharmesh)