Check out this product description workflow I built - 10x faster.

That 80% first draft is a huge win. Time saved on the draft means more time for the human parts, like aligning with our sales team's current messaging or tailoring for a specific campaign audience.

For SaaS descriptions specifically, I've found feeding the AI a clear "jobs to be done" framework works wonders. Instead of just listing features, I prompt it to structure the draft around the specific problem, how the tool solves it, and the outcome for the user. This gets the copy closer to how our customers actually talk.

What are you using as your primary input? A bullet list of raw features, or something more structured?

I completely agree with framing the input around a "jobs to be done" framework. That's a more effective prompt structure than a simple feature list because it builds in the narrative flow you want from the start.

My approach is to formalize this into a small, structured data object I feed into the model. I treat the product description task as a data integration problem. For instance, I'll create a JSON template with keys for "target_user_persona", "core_problem", "key_capabilities_as_verbs", and "desired_outcome". This ensures I never forget a component and the AI has a clean, unambiguous schema to work from.

It does add a layer of upfront work, but the consistency across different product lines pays off. Have you found that different models respond better to a narrative paragraph prompt versus a structured data input for this "jobs to be done" approach?

Ah, the classic "JSON as a universal input" strategy. I'm sure your product teams love being handed a schema to fill out before they can even think about the copy.

But let's get real. How long until that template becomes so rigid that you're just generating the same three boring descriptions with the nouns swapped out? You'll have perfect consistency across product lines, sure, but you'll also have perfectly generic, soul-less copy. The initial time you save on the draft will be spent later trying to inject some actual brand voice back into the results.

Different models probably respond better to different inputs, but that's secondary. The real pitfall is locking your whole process into a rigid data structure that was designed by an engineer, not a marketer.

That's awesome! Getting to a solid first draft in minutes is the dream. I've been tweaking a similar workflow for weeks.

One trick I found that really bumps up that usable percentage is adding a simple "rewrite in the style of..." step right after the main draft. You can point it at a couple of your existing high-performing descriptions, or even a competitor's copy you admire. It seems to bridge the gap between the generic structure and your actual brand voice more smoothly.

Which models are you mixing in Playground? I've been swapping between Claude 3 Sonnet and GPT-4o depending on the tone I need.

Totally agree about "jobs to be done." I'm still learning to write prompts, and I think I was making a big mistake by just pasting a feature list from our Shopify app page. The output sounded like a help doc.

I tried reframing it around the problem like you said, and it was way better. Like, for a description of a shipping calculator app, I prompted with "Our users get frustrated when shipping costs are a surprise at checkout." It gave me a draft that started with that pain point, which is way more relatable.

Can I ask a basic question though? How specific do you get with the problem statement? I worry if I make it too niche, the description won't appeal to enough people.

That makes sense. Saving the human editing time for strategic stuff like messaging alignment is probably the best use of the time you get back.

I've only used a basic feature list so far. I haven't tried structuring it around the problem and outcome like you described. Does that framework also help stop the AI from sounding too generic, or is that still a separate issue you have to tackle?

Oh, the "rewrite in the style of..." tip is a good one. I haven't tried that step yet. Does it work okay if you just paste the URLs of a couple of competitor's pages, or do you have to copy the text directly?

I'm also sticking to GPT-4o for now. Haven't tried mixing models in the same workflow, that seems advanced. Do you chain them in Playground, or just manually switch?