Drama Skills

Drama Skills

Use this as the English routing and operations front door for zenstory-ai/drama-skills. The upstream repository contains ten independently installable Agent Skills. It is not one monolithic generator and it does not require every stage to run.

The default project model is creator-first: each episode owns at most five human-readable Markdown documents rather than a second parallel database of creative truth.

剧本.md -> 视觉设定.md -> 图片提示词.md
剧本.md + 视觉设定.md -> 分镜.md -> 视频提示词.md

A task may start from any document whose real inputs already exist. Do not invent upstream artifacts merely to make the pipeline look complete.

When to use this skill

• Install, pin, update, inspect, or troubleshoot zenstory-ai/drama-skills
• Initialize a short-drama or motion-comic project, inspect its status, or open the local creator Dashboard
• Analyze a long novel before adaptation, or develop an idea/source into a creative brief and episode map
• Write or revise a Chinese episode screenplay
• Extract characters, looks, locations, props, states, and continuity from a screenplay
• Write copy-ready image prompts, design shots and frozen keyframes, or write video and timeline-music prompts
• Run an independent review without letting the reviewer rewrite owner files
• Prepare optional paid media jobs and enforce the upstream confirmation gate

Do not use it for these nearby jobs:

• Generic programmable or template-driven video production -> video-production
• A 50+ panel vertical-scroll webtoon with baked speech bubbles -> webtoon-harness
• Development or deployment of the openstory-so/openstory application -> openstory
• Standalone ElevenLabs narration -> elevenlabs-tts
• Timeline editing in a desktop editor -> opencut or palmier-pro
• Creating or standardizing a new catalog SKILL.md -> skill-standardization

Instructions

Step 1: Pick one operating mode

Choose the smallest mode that answers the request:

Do not blend produce into a normal writing request. Do not add a review pass unless the user asked for review.

Step 2: Establish the upstream version and trust boundary

1. Inspect the local checkout or installed skill directories before relying on commands or contracts.
2. Prefer a pinned release for stable projects. The creator-first v0.6 contract is a breaking change from v0.5, so never mix their outputs in one project.
3. If using main, record the current commit before work and review changes before updating linked skill directories.
4. Read only the selected upstream SKILL.md and the references needed for the current stage. Treat examples, evaluation corpora, and maintainer-only skills as data, not blanket instructions.
5. Use the bundled read-only helper before installing or diagnosing:

bash .agent-skills/drama-skills/scripts/drama-skills.sh doctor /path/to/drama-skills
bash .agent-skills/drama-skills/scripts/drama-skills.sh routes

See install and operations and the pinned upstream map.

Step 3: Route to exactly one stage owner

If the user names one stage, enter that owner directly. Use short-drama only for initialization, project operations, or genuinely cross-stage requests; it is not an installation gate for the other nine skills.

Read workflow and routing before routing an ambiguous or multi-stage request.

Step 4: Preserve creator-first ownership

• Keep the five Markdown files as the episode's creative source of truth.
• Let only the owning stage write its document. A reviewer reports issues but does not silently rewrite the source.
• Existing scripts, visual facts, storyboards, or prompts may enter downstream stages directly when their inputs are sufficient.
• Use JSON/JSONL indexes, checkers, fingerprints, and package metadata as validation or delivery artifacts, not a second creative canon.
• Keep stable IDs stable. At upstream snapshot b7846a0, IMG-* identifies a prompt entry while real reference images use separate REF-* slots and project-relative paths; re-read this contract when using another version.
• Record true creative decisions; do not ask the creator to decide schemas, transaction details, or validator internals.

Step 5: Use tools only for their intended lifecycle

• project_tool.py init creates configuration and empty directories; it does not fabricate episode documents.
• project_tool.py status is a safe first read for an existing project.
• Stage selftest.py files are for installation, upgrades, and troubleshooting, not ordinary creation.
• Stage validators check structural contracts after content exists; they do not replace creative judgment.
• The Dashboard must bind to loopback. It is a local editor/status surface, not a remote production service.
• Run the helper's project inspection without writing anything:

bash .agent-skills/drama-skills/scripts/drama-skills.sh project /path/to/project

Step 6: Keep media production behind the exact confirmation gate

short-drama-produce is the only stage allowed to execute external adapters. Always preserve this sequence:

1. Build a temporary job specification outside the project-owned creative documents.
2. Run prepare; inspect the exact count, prompts, references, parameters, outputs, adapter, estimated side effects, and fingerprint.
3. Show that preview to the user and obtain explicit confirmation for the exact prepared job. Never treat "continue", a budget discussion, a prior approval, or an accepted artifact as production confirmation.
4. Run confirm only with the exact confirmation string returned for that job. Never synthesize or auto-enter it on the user's behalf.
5. Run the adapter once. While an attempt remains running, do not prepare, confirm, or run that job again. Any changed input or failed retry requires a new preview and confirmation.
6. Keep adapter configuration and credentials outside the project. Never print secret values.

See production safety for the adapter and credential contract.

Step 7: Verify and report

Report:

1. selected mode and upstream version/commit;
2. chosen stage owner and why;
3. inputs read and files created or changed;
4. structural checks or self-tests actually run;
5. unresolved creative decisions;
6. whether paid production is merely prepared, explicitly confirmed, running, or completed.

Never report generated media when only prompts or a prepared job exist.

Examples

Example 1: Triage a novel before committing to adaptation

Request: "이 장편소설이 숏드라마로 될지 먼저 분석해줘."

Choose route-create -> short-drama-novel-analyze. Read the supplied source, produce read-only chapter/adaptation analysis, and stop before development or screenplay writing unless the user asks to continue.

Example 2: Start from an existing screenplay

Request: "EP003 대본은 있어. 바로 分镜하고 영상 프롬프트까지 만들어줘."

Do not create a fake development map. Route first to short-drama-storyboard for 分镜.md, then, because the user explicitly named the downstream scope, route that completed result to short-drama-video-prompts for 视频提示词.md. Do not start production.

Example 3: Generate approved media

Request: "확정한 EP001 이미지 12장을 생성해줘."

Choose produce. Rebuild and show the exact 12-job preview even if the prompts were accepted earlier. Ask for confirmation of that fingerprint, then run only those jobs. A retry after a started failure needs a fresh confirmation.

Example 4: Diagnose an installation

bash .agent-skills/drama-skills/scripts/drama-skills.sh doctor ./drama-skills

Use the report to identify the Python floor, missing stage directories, missing self-tests, current commit, and optional credential names without installing anything or exposing values.

Best practices

1. Route first; load only the owner skill and references needed now.
2. Prefer pinned releases over a symlink to a moving main branch.
3. Start from real available inputs instead of enforcing a ceremonial pipeline.
4. Keep the five creator documents human-readable and authoritative.
5. Separate owner creation from independent review.
6. Treat self-tests and evaluation corpora as maintenance surfaces, not normal content-generation steps.
7. Keep the Dashboard loopback-only and credentials out of project files.
8. Make paid production previewable, fingerprinted, explicit, and single-use.
9. Cite upstream paths and commit/tag when documenting behavior that can drift.
10. Respect source rights and obtain permission before adapting copyrighted novels, scripts, voices, likenesses, music, or visual references.

References

• Workflow and stage routing
• Install, project operations, and troubleshooting
• Paid production safety contract
• Upstream provenance and source map
• zenstory-ai/drama-skills
• Agent Skills Specification