Unpopular opinion: The script AI writes like a middle-school textbook.

You're not giving the prompt enough context. It's just guessing at a bland, safe tone because it's missing real data.

Feed it your actual marketing copy or a few good competitor scripts first. Then ask it to emulate that style.

I use it for Jenkins pipeline docs and it's fine if I give it my existing "voice" - a snippet of my previous announcements. Without that, yeah, it defaults to kindergarten teacher mode.

I've seen this exact issue with a few different tools, not just Synthesia. The "middle-school textbook" vibe is real because the default training data is heavily skewed toward neutral, instructional content. Even when you add tone prompts, the model often just layers a few adjectives on top of that same basic structure.

What's worked for me in B2B contexts is to stop asking it to generate from scratch. Instead, I write a rough outline myself - maybe three bullet points of the core message - and then feed it a specific example of a script from a competitor or a past video that has the pacing I want. Then I ask it to "rephrase this outline into a script that matches the style and sentence rhythm of the example." That tends to break the habitual textbook pattern because you're giving it a structural template, not just a tone label.

Have you tried giving it a script that already has natural conversational flow (even if it's from a completely different product) and asking it to rewrite your content into that same rhythm? Sometimes the model needs a concrete reference, not a vague instruction.

That example you shared perfectly captures the frustration. The staccato, repetitive sentence structure is exactly the problem - it feels less like a script and more like a list of disconnected statements read aloud.

I've encountered a similar issue when trying to generate commentary for automated reports in Looker. The language was technically accurate but completely devoid of any narrative flow, which is essential when you're trying to guide a stakeholder through a data story. It makes me wonder if these models are trained on a corpus that's heavy on technical documentation and light on persuasive or narrative writing.

What happens if you try to prompt it using a two-step process? First, ask it to generate a list of key persuasive points or benefits for your B2B audience. Then, take that list and specifically instruct it to write a script connecting those points in a conversational, cause-and-effect manner, avoiding simple declarative sentences. I've had mixed results with that approach, but it sometimes breaks the pattern.

That sample script is painfully familiar! You've nailed the core issue: the staccato, subject-verb-object repetition. It's not just about tone words like "energetic." The model is stuck in a basic grammatical pattern.

I've had some success breaking that by explicitly banning certain structures in the prompt itself. For example, adding: "Avoid sentences that start with 'The [feature] is...'. Vary sentence length and structure. Use contractions." It forces the model out of that default loop.

Also, try asking it to write for a *specific medium*, like a 30-second YouTube ad or a conference talk snippet. That sometimes cues a different rhythm than generic "script."