Prompt Engineering · AI Workflow
AI Script Prompt Engineering: How to Get Actually Good Scripts from AI
Most people give AI three-word prompts, get generic output, and blame the tool. That is like blaming a camera for a bad photo when you never learned to focus. The difference between a script you delete and a script you record is four prompt components most creators skip entirely.
Why Prompt Quality Is the Entire Game
We ran a controlled experiment: 200 prompts for the same YouTube topic — "how to build a morning routine" — with varying prompt quality. The output scripts were scored blind by three retention analysts who did not know which prompt produced which script. The scatter plot was ugly. The worst-performing scripts came almost exclusively from prompts under 50 words. The best-performing scripts came from prompts that were specific, constrained, and opinionated.
The delta between the worst prompt and the best prompt was 41 points on predicted retention. Same topic. Same AI model. Different instructions. The tool is not the bottleneck. The instructions are. Every creator who says "AI scripts are all the same" is, without realizing it, describing their own prompting habits — not the AI's capabilities.
The Astryx Prompt Framework: Four required components, one optional. (1) Content Type — what kind of video is this? (2) Format Constraint — what structure must it follow? (3) Audience Signal — who is watching and why? (4) Tone Directive — what voice should it use? Optional: (5) Creative Constraint — a specific rule that forces non-obvious output.
Component 1: Content Type — Stop Making the AI Guess
"Write a YouTube script about productivity" is useless. The AI has no idea if this is a 90-minute video essay, a 60-second Short, a tutorial with screen recordings, or a talking-head rant. Each format demands completely different pacing, structure, and tone. The AI defaults to a generic middle ground that works for none of them.
Specify the content type with precision. Not "tutorial" — "screen-recording tutorial with voiceover." Not "review" — "hands-on product review with B-roll of the device." Not "commentary" — "talking-head commentary with occasional graphics overlays." This is not pedantic. The format determines everything about the output. A screen-recording tutorial script reads completely differently from a talking-head commentary script. If you do not specify which one you are making, the AI splits the difference and you get mush.
Bonus: specify the video length. "8-10 minutes" vs "15-20 minutes" vs "under 60 seconds" produces fundamentally different script densities and pacing maps. Leave this out and the AI picks something arbitrary — probably 5-7 minutes, the statistical average of its training data, which may be wrong for your channel.
Component 2: Format Constraint — The Structure That Prevents Rambling
AI models love to ramble. Given freedom, they will write an essay, not a retention-optimized script. Format constraints are the cure. Tell the AI exactly what structure to follow. Not vague instructions like "have a good hook" — specific structural beats like "open with a surprising stat, follow with the problem it creates, then deliver three solutions in order of increasing counterintuitiveness."
The best format constraints are restrictive enough to prevent bad output but open enough to allow variation. Examples that work: "Use the Problem-Agitation-Solution-Proof structure." "Follow a Hook-Data-Reveal-Data-Reveal-Data-Reveal-CTA pattern." "Alternate between explanation and pattern interrupt every 45-60 seconds." The constraint gives the AI a skeleton. Without a skeleton, you get a blob.
This is the component that separates prompt engineers from prompt typers. Typers ask for output. Engineers specify the output's internal structure. The difference shows up in the second minute of every script — does it build momentum or wander off topic? Wandering correlates 0.81 with "no format constraint specified."
Component 3: Audience Signal — Who Is Actually Watching
The same topic written for a beginner sounds nothing like the same topic written for an expert. "iPhone tips" for a general audience: "Here is how to use Face ID." Same topic for power users: "Here is the automation workflow that Apple does not document." The AI cannot read your mind. If you do not tell it who is watching, it writes for the average person — which means every sentence assumes zero prior knowledge, and your actual audience tunes out.
Effective audience signals are specific: "For creators with 10K-50K subscribers who understand YouTube basics but struggle with retention." Not "for YouTubers." The more specific the audience signal, the more the AI calibrates depth, pacing, and assumed knowledge correctly. General audience signals produce general output. Specific audience signals produce output that feels like it was written for your exact viewer.
One trick that consistently improves output: tell the AI what the audience already knows. "My audience already understands SEO basics. Skip the definitions. Jump straight to advanced tactics." This prevents the AI from spending the first 90 seconds explaining concepts your viewers already understand — which is the #1 retention killer for established channels.
Component 4: Tone Directive — Kill the Wikipedia Voice
Default AI tone is academic-neutral. It hedges. It qualifies. It writes like it is afraid of being wrong. YouTube audiences punish this tone with 23% lower retention at the 30-second mark compared to confident, opinionated delivery. Your tone directive needs to override the default.
The most effective tone directives are persona-based, not adjective-based. "Write confidently" is useless — the AI does not know what confident writing looks like. "Write like a tech reviewer who has tested 200 phones and is tired of marketing lies" works because the persona comes with built-in voice patterns. The AI can model a persona. It cannot model an abstract adjective.
Specific tone directives that have performed well in our testing: "Write like a skeptical engineer," "Write like a coach who has seen too many people fail at this," "Write like someone who used to believe the opposite and changed their mind," "Write like you are annoyed that nobody talks about this honestly." Each persona produces a distinct voice. Pick one that matches your channel.
The optional fifth constraint — creative constraint — is the secret weapon. Add one specific rule that forces non-obvious output: "Do not use the word 'game-changer' anywhere in this script," "Every claim must be backed by a number," "The word 'actually' can only appear once." These small restrictions force the AI out of pattern-matching mode and into something closer to creative problem-solving. One constraint is good. Three constraints is too many — the AI starts contorting the content to satisfy rules and loses coherence.
The Prompt That Produced Our Highest-Scoring Script
For reference, here is a prompt that produced a retention score of 84/100 in our testing — the highest we have seen from a single prompt generation. Not because the prompt is magic, but because it specifies all four components precisely:
"Write a 12-minute talking-head video essay script about why most productivity advice fails. Format: Hook with a counterintuitive stat → three evidence-based counterpoints to common advice → personal failure example → synthesis conclusion with actionable takeaway. Audience: knowledge workers aged 25-40 who have tried GTD, Pomodoro, and bullet journaling and are frustrated. Tone: like a behavioral economist who has read the studies and is annoyed by the influencer version of productivity. Constraint: no sentence over 25 words."
Four components, one creative constraint, 97 words. Total time to write: under 60 seconds. Total time savings vs writing from scratch: hours. This is not advanced. It is just specific. Most prompts fail because they are vague. Be specific about what you want and the AI will give it to you. Be vague and you will get vague output. The prompt is the product.
Next Steps
Put prompt engineering into practice:
Tired of fighting with prompts to get usable scripts?
Astryx bakes prompt engineering into the generation engine. Tell it your topic and niche — it handles content type, format, audience calibration, and tone automatically. Then scores the result against real retention data.
Try Astryx Free →