Asking an LLM to write a short performance script sounds simple. Getting one that is coherent, playable, correctly structured, and suitable for a particular cast is a more demanding task.
For a project that generates scripts for live audio performances, I developed a staged pipeline: generate compact candidate story spines, score them, expand the selected candidate, generate the script, and validate it. Repair is bounded and should preserve the approved plan.
Explore alternatives before adding detail
The early stages produce several compact candidates rather than several complete scripts. A scoring rubric compares ideas before the application commits to expanding one.
That separation matters. If every candidate contains a full detailed plan, the exploration stage grows large and expensive. The retained notes describe a comedy iteration where over-detailed candidates contributed to output bloat and truncation; the correction was to restore compact candidates and expand only the winner.
Validate the output contract
The final script must satisfy requirements including cast mapping, narrator behavior, scene and word budgets, dialogue, and structured output.
These checks can catch an omitted character or invalid JSON. They cannot decide whether a scene is compelling. A script can pass every structural check while repeating the same emotional beat or introducing an unsupported revelation at the end.
The project therefore keeps creative review alongside deterministic validation. The notes identify failures such as static confrontation, late new rules, excessive narration, and endings that stop just as the premise becomes interesting.
Comedy exposed different weaknesses
A structure that supports suspense does not automatically produce comedy. The comedy experiments found premises that became dramedy, procedural arguments, or increasingly large spectacles without enough onstage cause.
The working direction became a playable single-game sketch, with narrow line-level punch-up after a coherent script already exists. That keeps a wording pass from becoming an uncontrolled plot rewrite.
These are observations from particular project iterations, not a general ranking of model providers or a benchmark of creative ability.
Repair should have a limited job
The pipeline includes a repair pass when validation fails. Its job is to correct a specific defect while retaining the selected plan, rather than asking for an entirely new script after every error.
That gives the system a clearer control loop. Exploration chooses a direction, expansion develops it, generation renders it into the required form, and repair handles bounded failures.
The broader lesson is useful beyond fiction: several well-defined stages can make a generative workflow easier to inspect than one increasingly overloaded prompt. Each stage needs a clear purpose, and a later stage should not silently undo the decisions an earlier stage established.