short-drama-knowhow is the upkeep skill of the Drama Skills suite. A maintainer starts it in a read-only, explicitly authorized session, and it studies short drama project material: premise, outline, script, visual setup, prompts, storyboard, edit notes and, where authorized, the finished media.
It then writes what it learned as public reference guides with their conditions and failure limits stated, and strips source traces first. It also downgrades, narrows or removes older rules when counterexamples or newer models show they are too strict. It is not meant for everyday writing or media generation. The skill’s text is in Chinese.
What You Get
- A workflow for reading a project in order, from premise to outline, script, visual setup, prompts, storyboard and edit notes
- Three reference files: research cards and coverage, synthesis and promotion of findings, and blind forward evaluation
- A rule-level scheme that sets how much proof a new or changed rule needs before it is merged
- A procedure for rewriting findings with fictional genres and characters so no source detail leaks into public guides
Common Use Cases
- A maintainer distills a lesson from a finished drama project into a public craft guide
- A maintainer tightens or adds a blocking rule and runs an anonymous blind comparison before merging
- A maintainer removes or narrows an old rule after a counterexample or a stronger video model shows it no longer holds
What Makes It Unique
Most skills in the suite help someone write or produce a drama. This one maintains the rules those skills follow. It ties each change to a rule level, from blocking rules that need a blind evaluation down to style options that only need an anonymization check, and it lets a rule be weakened or dropped on a counterexample without the heavy evaluation.
Under The Hood
The skill is a prompt only: a short SKILL.md plus three reference documents, with no scripts of its own. It tells the agent to read non-public project files only when the maintainer has granted read-only authorization in an isolated workspace, and to look at images, audio and video there to check real audio-visual facts. It starts no external media generation. It asks the agent to write drafts, anonymize them and run the repository regression tests, and to record large gate changes in the repository notes, so it can write files in the repository.