JianYing Editor Skill

JianYing Pro keeps a project as a folder on disk. draft_info.json holds the timeline, draft_meta_info.json the project metadata, and every file the skill imports is copied into that folder beside them. This repository is a Python wrapper around that folder, packaged as a skill an agent loads before it starts editing. JyProject creates or opens a draft, appends clips to named tracks, resolves effect, filter, transition and animation names through synonym tables, turns a script into a voiceover and lays a subtitle under each generated sentence for exactly as long as that sentence’s audio, and writes the keyframes behind zoom moves. Twelve CSV files index the app’s own cloud music, sound effects, video material, filters and transitions by id, so an agent can choose from JianYing’s library by Chinese name. Screen recording runs through ffmpeg with a Tk window that logs clicks; a browser animation becomes a clip through Playwright. Export is the one step that leaves the file system, and the one that differs by platform.

Who built itOne GitHub account, created in 2016, with 148 public repositories and 60 followers. The repository has 3,858 stars, 492 forks and 16 watchers. Its contributor table names three people besides the owner: @twodogegg for the macOS work, @shaozheliu for the media-missing fix on 5.9 and @Maxinsomnia for macOS draft support. 114 of the 117 commits on main are the owner’s.

1. 01

   The repository is a folder an agent reads, and much of it is data

   The skill is 184 files, and 47 of them sit under assets/: three bundles of JianYing’s own artist-effect assets, with an effect.prefab, a main.scene, materials, textures and GLSL shaders, plus an 8-file demo set led by a 3 MB test video. The other 137 files come to about 2.6 MB, and the largest of them is not a script. It is data/cloud_music_library.csv at 262 KB, an index of music ids JianYing’s cloud already holds. SKILL.md is 10 KB and is the first thing an agent reads; rules/ adds eleven files. GitHub counts the repository as Python first, HTML second and GLSL third, and the GLSL is all inside the effect bundles. The repository was created on 2026-01-24, and main carries 117 commits between 2026-01-27 and 2026-09-11, 114 of them from the owner’s account. VERSION reads 1.7.0 and the newest CHANGELOG.md entry is dated 2026-09-11. No release or tag has ever been published.

2. 02

   One JSON file is the whole interface, and someone else wrote the mapping

   A project is edited by composing draft_info.json before JianYing opens it. The mapping from that JSON to Python objects is not written here: scripts/vendor/pyJianYingDraft/ is a copy of GuanYixuan’s pyJianYingDraft under Apache-2.0, Copyright 2024, carrying its own LICENSE, and scripts/vendor/README.md lists the local patches the way that licence asks. Those patches are the reason for vendoring: the draft format was moved to the draft_info.json layout 5.9 and 6.x use, local_material_id is generated from the file name so clips stop going missing, and probing falls back to ffprobe when pymediainfo is absent. requirements.txt has no entry for it, so the skill runs from a clone. On top of that library scripts/jy_wrapper.py is 70 lines and assembles JyProject from four mixins. Its save() does more than write. It patches the material ids of cloud assets the draft refers to and forces adjustment nodes active, because a draft naming a cloud item without those fields opens with the item missing.

3. 03

   The app locks its own drafts, so the wrapper asks it to let go

   Resolution is decided at the first JyProject(...) call and defaults to 1920x1080, and the rules warn twice that a portrait project created with the defaults comes out with black bars. On Windows, JianYing holds the project folder open while that draft is on screen, and creation raises PermissionError. The wrapper retries three times, and between attempts it reaches into the running app through uiautomation, reads whether it is on the home page or the edit page, and sends it home before trying again. A draft folder missing its JSON is treated as corrupt and, with overwrite=True, deleted and rebuilt. Names are sanitised and every path is checked against the drafts root before anything is removed, which is what the v1.5.0 notes mean by blocking path traversal.

4. 04

   Voiceover is the app’s own service, called with keys read off the machine

   scripts/universal_tts.py does not use a cloud account of its own. It reads a device id out of JianYing’s local TTNet configuration and an iid out of the newest log files under the app’s user data, then opens a WebSocket to sami.bytedance.com with app_id 3704 and an app key that ships inside the source, under a User-Agent identifying itself as JianyingPro/5.9.0.11632. An environment variable can turn TLS verification off, and the script prints a warning when it does; by default verification is on, which the v1.5.0 notes list as a fix. When that call fails the fallback is edge-tts and Microsoft’s voices. add_narrated_subtitles splits a paragraph on Chinese punctuation, generates one audio file per sentence, and puts a text segment over it whose duration is that audio file’s own duration, so the subtitle cannot drift away from the voice. The backend that answered first is then locked for the rest of the paragraph: if it fails later the call raises, rather than finishing the video in a second voice.

5. 05

   Everything that is not a file write is a click or a browser

   tools/recording/recorder.py is a Tk window that drives ffmpeg, using gdigrab and dshow on Windows and avfoundation on macOS, while pynput writes mouse and keyboard events to a JSON file beside the recording. scripts/smart_zoomer.py reads those clicks back as keyframes: push in, hold, return. For effects the app does not ship, scripts/web_recorder.py starts a Playwright browser, records the page to video and waits for the animation to set window.animationFinished, with a 30-second ceiling. Export is where the tool leaves the file system entirely: scripts/auto_exporter.py clicks the app’s export dialog through Windows UI Automation. Called on macOS it returns exit code 2 with the message that the draft should be exported from JianYing by hand.

• Auto export is Windows only, and the README names JianYing 5.9 or older as the version it is stable on. Called on macOS, scripts/auto_exporter.py exits with code 2 and tells the caller to export by hand.
• The README rules three things out before anyone installs it: CapCut, the international build, is not supported at all, neither is the phone app, and the features that run on JianYing’s GPU, such as smart matting, beauty filters and speech-recognition subtitles, cannot be called from code.
• The app updates itself and there is no way to stop it, which is open issue 24. A newer build moves the controls the export automation clicks, so a JianYing update can break the last step with nothing in this repository changing.
• Six issues are open. One, filed 2026-09-16, reports seven places where a failure yields a finished video that is wrong instead of an error; another, filed 2026-09-22, asks media probing to degrade instead of aborting an import when ffprobe is missing.
• The recorder’s “自动生成智能草稿” button runs scripts/jy_wrapper.py apply-zoom with --name, --video and --json. jy_wrapper.py has no argument parsing; its __main__ block creates a project called Refactor_Test_Project and never reads those flags. The function that would apply the zoom, apply_smart_zoom in scripts/smart_zoomer.py, is imported by nothing and opens with a relative import that fails when the file is run on its own. scripts/smart_rough_cut.py imports an api_client module belonging to a different skill, and catches the failure with a warning that leaves both of its names set to None.