AI-Assisted Screenwriting: Common Mistakes In Storytelling Prompts And How To Avoid Them

Effective prompts & strategies to use ChatGPT in story development

In the past, I’ve shown examples of how to use ChatGPT to create scripts and demonstrated basic prompt techniques as well as advanced strategies for using AI-powered text generation in the context of story development.

In this article, I want to highlight what I think are the most common mistakes we make when using ChatGPT for script writing and pre-production. I hope this list is useful for anyone looking for a collection of basic examples of how we can overcome things that prevent us from making our prompts more effective. We will cover:

• missing context
• querying conventions
• restrictions without alternatives
• missing output structure
• missing segmentation

Common mistakes

Here’s a list of common mistakes to avoid to make your prompts more effective. Note that these are not stand-alone, ready-to-use prompts, but rather parts of a toolbox from which you can build your individual strategy (depending on your project and working style).

I will divide this into a section on building context and another on building structure (in reality, of course, the two sections and their various aspects overlap).

Building context

❌ Missing context:

Create ideas for a sci-fi thriller series

✅ Even in simple or spontaneous brainstorming sessions like the one above, we should always be guided by the principle of presenting context in a concrete, specific, and detailed way.

For example:

Create ideas for a gritty sci-fi thriller series set in Berlin and about an 
unsuccessful actor who struggles with painkiller addiction

For example, we would add information about the format (sci-fi thriller series), the style/tonality (gritty), and some narrative elements (e.g. locations, or even lead characters and inner/outer conflicts).

Such a prompt provides enough context to steer results in the right direction, but it has one major flaw: it is

❌ Querying conventions

To get away from clichés and genre conventions, we need to nudge the model in a direction where it can “explore” new patterns.

✅ To this end, we can instruct ChatGPT to focus on the unusual, the uncommon, or the highly unlikely (depending on where we want to go with our exploration).

For example, the prompt above could be expanded like this:

Create ideas for a gritty sci-fi thriller series set in Berlin and about an 
unsuccessful actor who struggles with painkiller addiction.

Brainstorm new angles or approaches and prioritize ideas that are uncommon 
or novel. Focus on unusual milieus and story worlds.

✅ One could also extend the basic prompt above with instructions on thematic threads or symbolic meaning:

Create ideas for a gritty sci-fi thriller series set in Berlin and about an 
unsuccessful actor who struggles with painkiller addiction.

Incorporate the themes "power of scientific knowledge vs. mythological beliefs"
and "father-and-son relationship".

✅ Or we could address style/tonality, format, or narrative elements. For example:

Create ideas for a gritty sci-fi thriller series set in Berlin and about an 
unsuccessful actor who struggles with painkiller addiction.

Balance intense action and suspenseful situations with light-hearted humor 
and wit.

If we take some time to carefully craft a prompt that contains useful semantic context, the results will be a lot better. However, we may want to exclude certain avenues of exploration in order to get closer to our desired results. In this case, we should avoid

❌ Restrictions without alternatives

It is often very effective to tell the model what it should not do. However, this can lead to vague/clichéd story elements or odd repetitions of the avoidance instructions.

✅ So instead, when using constraints (or “negative prompting”), we should always mention an “alternative pathway”. That is, we should not only tell the model what not to do, but also what to do instead. For example:

Create ideas for a gritty sci-fi thriller series set in Berlin and about an 
unsuccessful actor who struggles with painkiller addiction.

Do not use cyborgs or time travel, instead come up with interesting and 
unusual sci-fi tropes from the fields of biology, nano-tech, 
or cognitive science.

Building structure

❌ Missing specification about the structure of the output

Structuring the output can be fairly straightforward if we can draw on established patterns that are common enough to be included in the model’s training data.

✅ Story models are a good example of this. ChatGPT knows quite a few of them (even the more exotic ones), and together with advanced prompting strategies, these are powerful tools for prompt engineering:

Create ideas for a gritty sci-fi thriller series set in Berlin and about an 
unsuccessful actor who struggles with painkiller addiction.

Use the Save The Cat Beat Sheet.

✅ Another way to achieve structured output is so-called one-shot prompting. This time, we explicitly tell the model everything it needs to know about the structure it should use for the output:

Create ideas for a gritty sci-fi thriller series set in Berlin and about an 
unsuccessful actor who struggles with painkiller addiction.

For each idea, use this template:

Title: [story title]
Central conflict: [the central conflict of the main character]
Dilemma: [the dilemma the main character faces as a result of his or her 
reaction to the conflict]
Relationship line example: [main character's problematic relationship with 
another character]
Action line: [brief summary of the actions the main character takes to overcome their
problems to cope with their problems, eventually leading to an intensification 
of the original dilemma]
Twists: [possible twists during the climax]

❌ Missing segmentation

When you start combining the above methods for context creation and output structuring, you will eventually be dealing with complex prompts that consist of multiple segments, e.g., context, instructions, output structure, and so on.

✅ To keep these distinguishable and prevent the model from completing the context section and repeating structural terms rather than generating new output, it is often very helpful to include segmentation. That is, using titles and separators such as “###” or “ — -” to structure the prompt syntax and refer to its elements.

For example:

Here are some story ideas:
- Jonas Becker, a down-on-his-luck actor battling a painkiller addiction, 
begins to have strange visions of cyborgs in the gritty streets of Berlin. 
At first, he chalks it up to side effects from his heavy drug use, but the 
visions persist, becoming more detailed and frequent.
- Jonas lands a minor role in an upcoming sci-fi thriller about a dystopian 
future where man and machine are indistinguishably merged. Strangely, the plot 
of the movie reflects his own unsettling visions, causing him to question the 
origins of his hallucinations.

###

Use the above story ideas to create some characters for a suspenseful psychological 
thriller, where the bleak reality of contemporary Berlin intersects with a 
potentially hallucinatory cybernetic underworld.

The characters should explore themes of addiction, identity, 
and the blurry line between perception and reality.

For each character use this template:
[name, age]
[profession]
[biggest dream in life]
[character flaw that prevents them from living their dream]

With these common mistakes out of the way, I hope you’ll have fun developing stories with ChatGPT. I recommend you also take a look at more advanced concepts and prompting strategies to further improve your workflow:

➡️ For more information about AI & Creativity, follow me on Twitter or Medium (use my referral link to get full access to all my articles and those of thousands of other writers).

➡️ If you like my content, why not leave a “clap” at the end of this article, so more people can see it?