I’d like to preface with a disclaimer: I do think you should be using agentic tools and engaging with them. This is less a discussion about the tools themselves, and more about how you should go about using them (and a particularly sharp edge I see).
I’m writing this because slop is top of mind.
Yesterday I got a series of emails. The first one I received definitely had hallmarks of AI writing, but I was genuinely interested.
Then the second one came in. From a seemingly totally distinct sender, the content was distinctly similar, but the structure was identical. Surface-level insights from the show notes of my podcast. A mention of a recent LinkedIn post. The vaguely sycophantic tone. The verbosity.
The vague alignment, but fundamental disconnect and misalignment from what I’m trying to do, and why.
And then the third email rolled in. And the fourth. And fifth. And so on.
Defining Slop Link to heading
I have a trivial definition of slop: generated content that the author didn’t proofread. It’s apparently important enough to send, but not important for the author to actually read.
This disconnect / low-intent is what differentiates “slop” from useful generated content. I use agents to generate content for my consumption all the time! That isn’t slop - I can learn and benefit from it.
Slop is bad for you (the sender) Link to heading
Slop is bad in part because there’s so much of it.
You, a business owner, want to drive some outcome. Probably a sale, but maybe awareness. Maybe you want to get on my podcast. Maybe you just want to connect. Maybe you want a job, or to connect to do business.
Slop is like spam. To run a successful business, you need to get noticed and cut through the noise. Slop does the opposite - it slots you neatly into the center of noise.
We’re also kind of in a golden age of slop, just because spam filters haven’t caught up yet. But slop emails are being marked by spam, so bayesian filters are being trained to identify them. It’s a continuation of the same arms race, so inevitably outreach like this will start hitting your domain, and eventually these messages will just be lost in your spam filter like the rest of the financial scams, pill websites, and so on. Humans may be spotty on recognizing AI generated content, but bayesian filters won’t be spotty once trained.
Slop is just rude Link to heading
Slop sends a clear message: “My time is more valuable than yours”. It’s the opposite of leading with value, and signals that you don’t value me.
Why would I want to do business with you? You’ve already told me that you don’t value me as a customer. Why wouldn’t I go to a competitor who values more highly?
This also applies within a workplace. This immediately damages company culture, and sends a clear message that some employees are above other employees. It gets hackles up. Makes people defensive of their time. Speed is quick on a tactical level and slows on the strategic level. Work piles up. Communication breaks down. Politics takes hold.
Where to go from here? Link to heading
Well, there’s a simple solution - check your work, and reign in agents. Your agents should have to justify their skill level before you give them autonomy. As these are agents though, that’s work on you.
A good first step is prompting with your voice. An easy way to start is to have your agent go through your written materials (Emails, Linkedin, blogs, slack, etc), and summarize your voice. I also recommend custom parameters to decrease how “AI” it sounds - avoid verbosity, and prompt it to be concise or terse.
It’s also helpful to have a “AI disclosure”. If the email is sent by “Clippy, X’s agentic helper”, you’ll do some damage to your personal brand, but damage will be concentrated around the agentic technology, not you. This is weird and subjective, I know, but knowing content is AI generated means I’ll give it grace, when otherwise I might feel like it’s pretending to be human. It’s a form of responsibility laundering, but better than nothing. Worst case “Sent with Clawd” or “Sent with PodPitch” are helpful in separating you from the technology. If it’s coming in from you, you (the individual) are assuming full responsibility.
Next, have it draft multiple drafts before sending or publishing (to yourself). This is especially important when you’re starting - you can see patterns in how it writes. Think to yourself - what if my customers get multiple messages each with this pattern? Human communication is messy, not over-formatted. An option may be to increase the temperature of the model you’re using, if you have access to that config lever.
Another useful experiment is to run your agent across multiple models. This is an insurance policy for you. A great prompt will work across Haiku, Sonnet, Opus, and ChatGPT. If you’ve completed steps above and you’re getting more varied (good) content, you can now more freely swap between models without causing a regressions in quality.
Next, start preparing emails for real prospects. Depending on your agent, have them write the emails in draft, and send them yourself. You can edit before send, but don’t forget to iterate on your prompt & context.
There’s further optimization to be done after this, but I imagine this is a large workflow for many individuals to follow already. If you’ve worked in Engineering or Quality, this will look very familiar. Keep in mind that these tools still have a significant defect rate, and that those defects must be managed.
That is, if you want outcomes from using these tools in the first place.