Create Story YAML

Create a gptauthor story YAML

Expand a short story idea (a sentence, a paragraph, a premise) into a complete gptauthor story prompt file, then validate it so gptauthor --story <name> runs without errors.

The human writes the plot; gptauthor writes the prose. This file is the plot — it must contain the whole arc including the ending. Do not leave the ending for the LLM to invent.

Workflow

1. Read the user's short story text and note which decisions it already settles.
2. Ask the user about the style with a single AskUserQuestion call, before writing anything (see Asking about style).
3. Pick the output path (see Filename and location).
4. Copy assets/template.yaml and fill it in, applying the answers (see Writing the content).
5. Validate: python3 .claude/skills/create-story-yaml/scripts/validate_story_yaml.py <path> Fix every ERROR; treat WARNs as a review checklist.
6. Report the path, how the answers shaped the file, and the run command.

Asking about style

A short premise never settles audience, tone, focus, or ending, and every one of those changes the whole file. Ask once, up front, in one AskUserQuestion call — do not write a draft first and revise it afterwards.

Rules:

• Skip anything the user already specified. If the premise says "a kids' book", don't ask about audience. Never ask a question the prompt already answers.
• Four questions maximum per call. Pick the four most open dimensions from the table below; if fewer than four are open, ask fewer.
• Put the option you'd recommend first and mark it (Recommended). The user always gets a free-text "Other", which is where the interesting answers usually come from.
• If the user declines, answers "you choose", or is clearly in a hurry, pick sensible defaults, write the file, and state each choice you made in your reply.

Other dimensions worth a slot when they are the open ones: comparable authors to write toward, real people or events to stay faithful to, story length (a 3-chapter run wants ~8 beats, a 15-chapter run wants ~15), and anything the story must avoid.

The most valuable answer is usually free text on key story points to hit — the user's own beats. Invite it explicitly in the Focus question's phrasing, e.g. "Where should the story's weight sit? (Pick any, or use Other to list specific beats you want hit.)"

Then use the answers:

• Audience sets vocabulary, sentence length, chapter length, and what content is on the table. Say it plainly in the style block, e.g. The book is written for readers aged 8 to 12: clear sentences, concrete images, no graphic violence.
• Tone picks the comparable authors and the list of story devices.
• Focus decides which beats get several lines and which get one.
• Ending goes in the beats and, when it needs pinning down, a Final chapter: block.

Required schema

All five top-level keys are required, spelled exactly like this. Nothing else is read.

system is a plain one-line string; prompt is a block string |-. A missing system or prompt key is an assertion failure at runtime.

Placeholders

prompt values go through Python str.format(). Only these names are supplied — anything else raises KeyError mid-run:

• synopsis.prompt — {total_chapters} {book_description} {book_characters}
• expand-chapter-first.prompt — the above plus {synopsis_response}
• expand-chapter-next.prompt — the above plus {previous_chapter_number} {previous_chapter_text} {chapter_number}

Rules that follow from that:

• Literal braces in a prompt must be doubled: {{ and }}.
• system is never formatted. Placeholders there are passed to the model verbatim.
• expand-chapter-next.system is dead — the engine uses expand-chapter-first.system for every chapter. Keep the two identical so a later edit isn't silently ignored.
• Never write ${...} anywhere in the file; OmegaConf treats it as interpolation and the run dies on load.
• Never hardcode a chapter count in prose — the count comes from --total-chapters. Always use {total_chapters}.

Keep the three prompt bodies close to the existing files (gptauthor/prompts-*.yaml); they are tuned to parse correctly. Vary the system line and the genre nouns to match the story, and leave the mechanics alone. assets/template.yaml already has working versions.