lazyedit-publish-workflow
Use when publishing videos through LazyEdit, AutoPubMonitor, a remote AutoPublish host, Shipinhao, YouTube, Instagram, or LALACHAN-generated videos; covers direct CLI/API publishing, current-run reuse
SKILL.mdskills/lazyedit-publish-workflow/SKILL.md
GitHubLazyEdit Publish Workflow
Use this skill for normal LazyEdit publish tasks and for AI-generated videos from LALACHAN/RARACHAN that need subtitle correction, processing, and platform publishing.
Runtime Map
- LazyEdit repo/backend:
$LAZYEDIT_ROOT - Studio app:
$LAZYEDIT_STUDIO - LazyEdit API:
$LAZYEDIT_API - Publish CLI:
scripts/lazyedit_publish.py - AutoPubMonitor repo:
$AUTOPUB_MONITOR_ROOT - Nutstore import folder:
$NUTSTORE_AUTOPUBLISH - Remote AutoPublish host:
ssh $AUTOPUBLISH_SSH - Remote AutoPublish repo:
$AUTOPUBLISH_ROOT - Remote publish API:
$AUTOPUBLISH_API - Remote tmux session:
autopub
Keep local paths outside git. From the LazySkills repo, copy the example config, edit it for the machine, then source it before using command examples:
cp .config/lazyskills.env.example .config/lazyskills.local.env
$EDITOR .config/lazyskills.local.env
set -a
. .config/lazyskills.local.env
set +a
Hosted editing and publication capability
For the invite-only Studio service, editing and preview are the normal member
workflow. Read authenticated /accounts/account and /auth/me capabilities;
publication requires capabilities.publishing=true and the granted
publication.publish scope. A disabled account uses prepare/process/preview,
not a social login desktop or publish call. Only the operator's protected
backend policy enables publication for an immutable account ID; invitations,
subscriptions and OAuth login do not enable it. The reviewer has ordinary member
capabilities and must never receive the personal Pi, profiles, library or channel
credentials. Use the account's scoped API base and private client state.
See LazyEdit's references/studio/editor-first-and-optional-publication.md for
operator configuration and linked-app handoff. This does not change the original
owner's local/Pi workflow or authorize a new real post.
Core Rule
Prefer the LazyEdit CLI over manual browser work. It creates normal LazyEdit jobs, so the webapp queue stays in sync.
For a normal publish, run one lazyedit_publish.py command and give LazyEdit a
reviewed story/script context file with --prompt-file. LazyEdit owns subtitle
correction, metadata generation, rendering, packaging, queue submission, and
monitoring. Do not pre-write, replace, or directly package metadata merely to
make it look better. Separate --correction-prompt-file,
--metadata-prompt-file, or manual metadata JSON are recovery tools only after
an inspected normal run proves that the shared context cannot produce correct
results.
When confirmed speech conflicts with ASR, Korean restoration needs review, or a native submission failed before dispatch, read context correction and native recovery.
Activate the environment first:
cd "$LAZYEDIT_ROOT"
source ~/miniconda3/etc/profile.d/conda.sh
conda activate lazyedit
Safety Rules
Published staging cleanup is enabled by default in AutoPublish. Let
published_retention.pyverify archive hashes and every requested platform result, then remove only archived media/ZIP copies. Preserve original media, Nutstore masters, corrected lyrics, covers, proofs and receipts. Failed, active, ambiguous or test jobs are not cleanup candidates. For a deliberate later republish, re-upload the retained canonical ZIP instead of assuming the remote staging copy still exists.If Shipinhao stays at cover
生成中or music upload stays at 0%, inspect resource timing before blaming disk space or changing video resolution. AutoPublish'sreferences/shipinhao-dependencies-and-published-retention.mdcovers verified public WASM caching and the asynchronous language-dropdown fix. Never bypass readiness or invent successful publication receipts.Music submission needs native form validation, not merely an enabled button. Wait for audio/cover uploads and all form validators; require the explicit submitted page. If a submit click has uncertain results, inspect management before retrying. Include the public source URL in the music package and fill the external playback field when it was already released. Keep production autoreload off and deploy only to an idle queue.
Do not publish to real platforms just to debug packaging, subtitles, or logo output. Use
--no-publishfirst, inspect the generated ZIP/final MP4, then publish exactly once when the package is correct.For current Musia recording videos, publish with the existing LazyEdit logo at top-right and no LazyEdit subtitles unless the user explicitly asks for subtitles. Use
--no-burn-subtitles --logo --logo-position top-right, force a fresh logo-only render when the side changes, and inspect a sample frame or the MP4 inside the ZIP before submitting.For native portrait Musia player recordings, also pass
--no-portrait-blur-fill. Do not apply portrait bg-fill to a video that is already a direct 2160x3840 portrait capture. Bg-fill is only for converting horizontal or nonmatching-aspect source videos, and the foreground geometry must respect the source aspect ratio.If a Musia recording already has the correct logo burned into an inspected high-quality MP4, do not let LazyEdit reprocess it. Build a direct AutoPublish ZIP with a unique stem, matching
<stem>_metadata.json,video_filename, andcover_filename, then POST it to/publishwith the requested platform flags. This is the preferred recovery path when a LazyEdit--no-processroute still triggers processing or overwrites the inspected logo master.For 4K Musia player recordings with small text, compare source and processed outputs before publish. If a logo-only/subtitle-disabled LazyEdit render is much smaller, lower bitrate, or visibly blurrier than the approved source, preserve quality by direct-packaging the inspected MP4 for AutoPublish. Keep a sample-frame check beside the package when debugging.
For Musia posts, platform metadata must describe the song itself, not the workflow. Do not include conversation context, model/pipeline notes, recording mode names, "publication master", "no subtitle", implementation details, or internal review notes unless the user explicitly asks for a technical demo. Good metadata focuses on song title, artist, theme, mood, language mix, genre, and a short listener-facing hook.
If generated metadata invents languages, wrong lyric phrases, model names, or private workflow context, first correct the context supplied to LazyEdit and rerun its normal metadata step. Use manual metadata JSON only as an explicit recovery path when the normal generator remains defective; record the failure instead of silently making manual metadata the default.
Real publishes should use polished/corrected subtitles and the configured LazyEdit Studio logo unless the user explicitly asks otherwise. Verify logo settings with
curl -fsS $LAZYEDIT_API/api/ui-settings/logo_settings | jq .; normal logo outputs end in_subtitles_logo.mp4.For Musia pure-music publishes, lyrics must come from the corrected Musia website/publish lyric JSON for the exact selected vocal, not from the original prompt lyric, draft lyric, or master-language lyric file. If only a draft is available, stop and run the Musia ASR/listening correction workflow before creating a Shipinhao Music or YouTube Music package.
For LALACHAN/RARACHAN generated videos, pass a reviewed story/prompt context with
--prompt-file. It is evidence for both subtitle correction and metadata, not a verbatim transcript or metadata body. Include correct character names, setting, central action, tone, and payoff; omit local paths, production chatter, and unrelated instructions. LazyEdit must derive concise viewer-facing metadata itself.Do not manually rewrite or replace metadata during a normal publish. If the generated result becomes a storyboard dump, improve the supplied context or report/fix the LazyEdit metadata behavior. Split prompt files only for a diagnosed exceptional case.
Before every generated-video burn or publish, compare the source ASR SRT with the polished SRT using
scripts/validate_subtitle_timeline.py. Ordinary text correction must preserve cue count and every cue's start/end timestamps. A failure blocks burn and publish. Never let a story script replace dialogue in sequence while reusing unrelated ASR timestamps.If correction is expected to recover dialogue that ASR missed, inspect
DATA/VIDEO_FOLDER/*_mixed_polished.mdand align the added line against the actual audio as a separate operation. Do not insert it by shifting later text into earlier cues. After validation, spot-check audible dialogue near the start, middle, and end before any platform post.If missing-language recovery creates plain subtitle text, do not restore grammar colors with a per-video patch. Fix or use the shared
lazyedit/subtitle_tokens.pynormalization path so plain text, ruby markup,word/readingtokens, and speaker-helper rows all render through grammar-typed palette tokens.When copying through Nutstore, use one stable
_COMPLETEDfilename and watch AutoPubMonitor panes before recopying. Avoid creating duplicate source files just to retrigger the watcher.When the source path is already under LazyEdit
DATA/, use--video-idor a non-colliding--filename. Do not re-uploadDATA/<stem>/<filename>with the same filename, because the upload endpoint can truncate the source by writing over it.For XiaoHongShu, close hashtag suggestion popovers with Escape/blur before the final publish click. The red publish control may be inside a custom
xhs-publish-btn, so use the AutoPublish fallback instead of hand-clicking random page coordinates.For Douyin, reuse an existing unpublished draft when the upload already exists. Do not use native
send_keys()for title/description fields or the separate topic widget when debugging; the site can wedge Selenium. Use the AutoPublish JS field replacement path and keep hashtags in the description.For Bilibili, optional SMS verification after upload is only for completion notifications; close it and continue. If the page shows
请完成短信验证while the upload is stuck at0.0MB/0.0MB, it is a hard SMS gate, not GeeTest. Click获取验证码, get the SMS code from the user, and do not retry upload loops without it. Cover upload is best-effort, so a missing cover dialog should not cause a full reupload.If Bilibili shows
0.0MB/0.0MBand browser-sidepreuploadreturns code601with您上传视频过快,请您稍作休息后再继续, stop retrying and wait for cooldown. Repeated upload retries extend the block.To add a missing platform to an already-processed LazyEdit output, reuse the existing ZIP if it contains the correct rendered MP4. Re-submit the same ZIP with only the missing platform flags. Repackage only when the existing ZIP points at the wrong output.
For
scripts/lazyedit_publish.py --platforms, use the CLI's accepted video platform names:shipinhao,instagram,youtube,douyin. Do not pass shorthand aliases such assph,ins,y2bto this CLI; they can fail after processing. Shorthands may still appear in lower-level AutoPublish recovery flags, so distinguish the LazyEdit CLI layer from remote recovery/debug endpoints.For normal public LALACHAN story/video publishes, include Shipinhao by default:
--platforms shipinhao,youtube,instagram,douyin, unless the user explicitly excludes it. If Shipinhao login blocks a combined job, publish the remaining platforms separately only as a partial fallback, then leave an explicit Shipinhao-only retry using the same verified ZIP. Do not report the overall publish as fully complete while Shipinhao is pending.If Shipinhao is logged out, the expected behavior is to extract the QR code and send the login email. If AutoPublish fails earlier with a message such as
Shipinhao login iframe was not available and the publish editor is not ready, treat that as a Shipinhao login-flow bug to fix or report. Do not keep restarting all platforms; preserve the ZIP and retry Shipinhao only after the login/QR path is corrected.AutoPublish derives the extracted metadata directory from the ZIP filename stem. Do not rename a prepared ZIP to add suffixes like
-toprightunless the internal metadata filename and directory contract are regenerated to match.If one platform browser looks stale after an interrupted publish, use the AutoPublish per-job recovery controls instead of globally restarting every browser: POST the same ZIP to
/publishwith the normal platform flags plusrestart_platforms=ins,y2b,sph,douyinfor the affected platforms, orforce_browser_restart=truewhen all selected target browsers should be restarted. Check/publish/queue; queue rows now include these restart flags.When posting a direct AutoPublish ZIP to the remote API, use
filename=<zip-name>plus boolean flags such aspublish_shipinhao=true,publish_y2b=true,publish_instagram=true, andpublish_douyin=true. Do not usepath=...for this endpoint. AutoPublish expects the ZIP under itstranscription_data/<stem>/directory and derives the metadata directory from the ZIP stem.If the AutoPublish server or tmux session is interrupted mid-run, its in-memory queue may be lost. Requeue the same verified ZIP rather than rebuilding. If a platform may already have clicked publish, recover the remaining platforms first and only retry the uncertain platform after checking drafts/history.
Avoid opening many long-lived terminal monitors. Prefer one
scripts/lazyedit_publish.py --guided-monitor --waitprocess plus occasional one-shot queue/tmux checks. Close stale sessions before starting another long publish.
Setting Semantics
--use-current-settingsreads Studio defaults.- One-shot flags such as
--platforms,--languages,--subtitle-lift-ratio, and--no-burn-subtitlesdo not change Studio settings. - Only
--persist-settingswrites CLI options back to the webapp preferences. --languagesis bottom-to-top subtitle order.- For a destination-language travel episode, keep the established upper rows and place the local language last in the visible stack. For example,
--languages fr,zh-Hant,ja,enrenders top-to-bottom as English, Japanese, Chinese, French. Mark each subtitle item with its actual spoken language so the speaker/sound icon follows the source-language row instead of a fixed row. - Four-language vertical placement has two accepted one-shot styles. A lifted
band uses
--subtitle-lift-ratio 0.1; a bottom-anchored band uses--subtitle-lift-ratio 0. The latter keeps upper rows visually closer to an unlifted three-language layout when a bottom language is added. If the user does not choose a style, omit this flag and inherit Studio settings. Do not persist either style merely because one publication used it. - Reuse the AI provider already configured by LazyEdit. DeepSeek is the normal default on an existing DeepSeek installation; do not require an OpenAI key or create an OpenAI client before the selected provider handles translation or metadata.
- If Studio logo settings are enabled,
--no-burn-subtitlesstill creates a processed logo-only output ending in_logo.mp4and publishes that output. Translation is skipped because subtitles are disabled. - Use polished/corrected subtitles for real publishes and debug publishes unless the user explicitly requests original subtitles.
- Publish category defaults: personal phone/self recordings use
simplelife; LazyingArt brand/product posts uselazyingart; pure music/art-track posts usemusia; LALACHAN story videos uselalachan; LALACHAN character music videos uselalamv. Instagram has no stable per-post category/playlist in the desktop web upload flow, so AutoPublish only logs the inferred category there and uses normal captions/tags. LazyEdit metadata generation asks the model forpublish_category(simplelife,lazyingart,musia,lalachan, orlalamv) and the router falls back to source-path/keyword inference.musicis only a backwards-compatible alias formusia. Use--publish-category lalamv,--youtube-playlist LalaMV, or--shipinhao-collection LalaMVfor MV overrides. - Burn the existing LazyEdit webapp logo on real publishes unless the user explicitly says no logo. Use the configured Studio logo; do not upload or invent a new asset.
- Required logo state is
enabled: true,logoPathpresent, and the requested position set. Current Musia/LALACHAN/MV default isposition: "top-right". Check it before CLI/API publishes withcurl -fsS $LAZYEDIT_API/api/ui-settings/logo_settings | jq .. For no-subtitle logo-only publishes, force a fresh burn when the position changes, extract a sample frame withffmpeg, and visually verify the logo side before submitting to AutoPublish. --no-processreuses an already completed output. Use it when the user says "last run", "same version", or "already finished run".--publication-session-id IDtargets a specific run. Omit it for the current output.
2026-06-30 Musia/Platform Runbook
See $LAZYEDIT_ROOT/references/PUBLISH_RUNBOOK_MUSIA_AND_PLATFORM_SMOOTHING_2026_06_30.md for the full incident record. Key rules:
- Confirm the MP4 inside the ZIP is the desired version before real publish.
- Musia recordings: no LazyEdit subtitles by default; top-right logo; category
musia; use curated song context for metadata. - LALACHAN/generated story videos: use the full script for subtitle correction, but use a short metadata brief so public descriptions do not become storyboard dumps.
- Shipinhao collection and YouTube playlist selection are best-effort; if the requested collection/playlist is absent or not immediately selectable, publish and report the fallback.
- Douyin should reuse drafts and use the JS field replacement path.
- Xiaohongshu needs popovers closed before final publish.
- Bilibili upload cooldown/SMS gates should stop retries; do not solve SMS gates with GeeTest/Tuling.
- Shipinhao Music: publish as music/song, not album; use square covers; confirm cover overlays; fill lyrics, story/
音乐人说, language, genre, author, originality/agreement when visible; use corrected Musia website vocal JSON for the exact selected audio. Before posting, inspect*_lyrics.txtin the package and compare it withwebsite/data/songs/<song>/lyrics/<vocal>/<lang>.json, not the original prompt lyric or another vocal's translation JSON. Upload plain corrected lyric lines to Shipinhao Music; do not upload[mm:ss]LRC timestamps because the current desktop form rejects them as表单信息不完整. If proof upload remains at0%, wait, then remove the stuck optional proof instead of submitting a half-uploaded proof. If the live language dropdown lacks English/Japanese, publish and report the fallback instead of silently claiming it was selected. Usescripts/lazyedit_music_package.py --postwithout an explicit endpoint when possible; the CLI now probes reachable AutoPublish/publishendpoints. If overriding manually, pass/publish, not/publish/queue, because the queue URL is read-only and returns 405 to POST. - Music monetization beyond Shipinhao should start with a distributor-ready
bundle, not a brittle direct uploader. Use
scripts/lazyedit_music_distribution_bundle.py --package-dir DATA/music_publish/<slug>to export audio, optional WAV derivative, square cover, corrected lyrics, JSON/CSV metadata, and upload notes for Bandcamp, SoundCloud, or distributors such as DistroKid, TuneCore, CD Baby, RouteNote, or SoundOn. For official YouTube Music/Spotify/Apple Music delivery, use a distributor; LazyEdit's YouTube art-track MP4 is normal YouTube publishing, not DSP delivery. - Bandcamp music publishing is prepared but guarded. Use
--bandcamp-audiowith a real WAV/AIFF/FLAC master; never use an MP3-derived WAV as the Bandcamp source. First run with--platforms bandcamp_music --post --testso AutoPublish opens the port-5008 Bandcamp browser for registration/login and form inspection. SetBANDCAMP_UPLOAD_URLif the account has a specific artist upload page. LeaveBANDCAMP_PUBLISH_PUBLICunset until price, payment, metadata, and account settings have been manually verified.
Category Cleanup
Use platform cleanup scripts only after an inventory or dry run. They attach to
the logged-in browser sessions and are read-only until --apply.
Shipinhao collections:
ssh lachlan@lazyingart 'cd ~/Projects/autopub && /home/lachlan/venvs/autopub/bin/python scripts/manage_shipinhao_videos.py ensure-collection --collection LalaMV --apply'
ssh lachlan@lazyingart 'cd ~/Projects/autopub && /home/lachlan/venvs/autopub/bin/python scripts/manage_shipinhao_videos.py ensure-collection --collection Musia --apply'
ssh lachlan@lazyingart 'cd ~/Projects/autopub && /home/lachlan/venvs/autopub/bin/python scripts/manage_shipinhao_videos.py ensure-collection --collection 啦啦侠 --apply'
ssh lachlan@lazyingart 'cd ~/Projects/autopub && /home/lachlan/venvs/autopub/bin/python scripts/manage_shipinhao_videos.py inventory --scrolls 5 --output /tmp/shipinhao_inventory.json'
ssh lachlan@lazyingart 'cd ~/Projects/autopub && /home/lachlan/venvs/autopub/bin/python scripts/manage_shipinhao_videos.py move-category --category lalamv --lalamv-collection LalaMV --scrolls 5 --output /tmp/shipinhao_lalamv_plan.json'
ssh lachlan@lazyingart 'cd ~/Projects/autopub && /home/lachlan/venvs/autopub/bin/python scripts/manage_shipinhao_videos.py move-classified --scrolls 5 --output /tmp/shipinhao_move_plan.json'
Apply in small batches or by exact visible title fragment:
ssh lachlan@lazyingart 'cd ~/Projects/autopub && /home/lachlan/venvs/autopub/bin/python scripts/manage_shipinhao_videos.py move --query "visible title fragment" --collection 啦啦侠 --apply'
Shipinhao mirrored metadata/description management:
python AutoPublish/scripts/shipinhao_mirror_manager.py export-metadata --metadata-root DATA --days 45 --output /tmp/lazyedit_shipinhao_metadata_index.json
python AutoPublish/scripts/shipinhao_mirror_manager.py export-publish-history --limit 500 --output /tmp/lazyedit_shipinhao_publish_history.json
ssh lachlan@lazyingart 'cd ~/Projects/autopub && /home/lachlan/venvs/autopub/bin/python scripts/shipinhao_mirror_manager.py mirror --scrolls 5 --output /tmp/shipinhao_mirror.json'
python AutoPublish/scripts/shipinhao_mirror_manager.py sync-db --db /tmp/shipinhao_management.sqlite --mirror /tmp/shipinhao_mirror.json --metadata-index /tmp/lazyedit_shipinhao_metadata_index.json --publish-history /tmp/lazyedit_shipinhao_publish_history.json --output-plan /tmp/shipinhao_description_plan.json
python AutoPublish/scripts/shipinhao_mirror_manager.py db-report --db /tmp/shipinhao_management.sqlite --limit 20
Use this mirror manager for existing-post control, not publication. On
2026-06-29, old date-only rows could be matched back to LazyEdit metadata, but
Shipinhao's 修改描述和封面 page only allowed modifying selected existing text
with a 20-character limit. Blank/missing descriptions could not be restored
through the visible desktop UI; the tool reports
unsupported-description-repair for that state and records apply attempts in
the SQLite mirror DB. Inspect every JSON plan before --apply.
YouTube playlists:
ssh lachlan@lazyingart 'cd ~/Projects/autopub && /home/lachlan/venvs/autopub/bin/python scripts/manage_y2b_videos.py move-category --category lalamv --lalamv-playlist LalaMV --scrolls 20 --output /tmp/youtube_lalamv_plan.json'
ssh lachlan@lazyingart 'cd ~/Projects/autopub && /home/lachlan/venvs/autopub/bin/python scripts/manage_y2b_videos.py move-classified --scrolls 20 --output /tmp/youtube_move_plan.json'
For lalamv, target YouTube playlist LalaMV. Playlist selection and creation are best effort: if YouTube creates the playlist but does not expose it immediately for selection, continue the publish and report the limitation instead of failing the whole job.
Instagram:
Instagram does not have a comparable per-post category/playlist/collection target in the current desktop web upload flow. Do not run an Instagram category backfill. Keep using metadata category for YouTube/Shipinhao routing and normal Instagram caption/tags.
Never bulk-apply a generated plan without inspecting the JSON. If the page is logged out, wrong, or the visible row text is weak, stop and open the correct management page in the browser first.
Quality-Preserving Portrait Publish Masters
When a 4:3 or horizontal generated video is converted to vertical mobile format, do not accept visible quality loss from repeated normal delivery encodes. Keep the foreground sharp, use a blurred current-frame fill, and leave the lower blurred area for subtitles.
Preferred current path: use LazyEdit's built-in portrait blur-fill feature and normal subtitle/logo reburn. In LALACHAN mode, LazyEdit targets a lower blurred reserve of about 40% (bottomSpaceRatio=0.4) and derives the top margin from the source aspect ratio. Manual blur-fill scripts are fallback/recovery tools for layout experiments or older runs.
Recommended portable pattern from the LALACHAN repo:
cd "$LALACHAN_ROOT"
scripts/portrait_blurfill_subtitle_space.sh INPUT.mp4 OUTPUT_portrait_hq.mp4 \
--fg-y 544 \
--crf 10 \
--preset slow \
--scale-flags lanczos \
--audio-mode copy
For 1080x1920 portrait output from a 16:9 MV, the built-in bottomSpaceRatio=0.4 layout is preferred. If falling back to the manual script, --fg-y 544 roughly matches the same lower-space target for 16:9 material; avoid the older high-foreground --fg-y 240 unless deliberately reserving a much larger lower area.
In the LazyEdit Publish tab, use the Calculated layout card and View layout
modal to inspect source-specific top/foreground/bottom geometry before
processing.
If LazyEdit's normal subtitle/logo burn visibly reduces quality, create an already-burned high-quality publish master locally. Keep the normal configured logo position and style, usually top-right for current LALACHAN/MV work, and place subtitles mostly in the lower blur area:
scripts/hq_subtitle_logo_master.sh OUTPUT_portrait_hq.mp4 corrected.srt OUTPUT_publish_hq.mp4 \
--logo "$LAZYEDIT_LOGO_PATH" \
--logo-height 288 \
--logo-x 38 \
--logo-y 38 \
--font-size 44 \
--margin-v 280 \
--crf 10 \
--preset slow \
--audio-mode copy
Then import/publish OUTPUT_publish_hq.mp4 with --no-burn-subtitles. This avoids a second burn/re-encode because the MP4 already includes the normal logo and burned subtitles:
python scripts/lazyedit_publish.py \
--video-id VIDEO_ID \
--use-current-settings \
--platforms youtube,instagram,shipinhao \
--prompt-file $LALACHAN_ROOT/references/prompts/PROMPT.md \
--no-burn-subtitles \
--no-process \
--guided-monitor \
--wait
Use this path only after verifying sample frames prove the subtitles and logo are already present in the MP4. If they are not present, use the normal LazyEdit burn path.
Authoritative SRT Fast Path
When a reviewed SRT already matches the final video, import it directly instead of running Whisper again. This remains the normal LazyEdit pipeline after subtitle import: multilingual translation, ruby/pinyin rendering, blur-fill, logo, metadata, packaging, and AutoPublish all use the existing implementation.
python scripts/lazyedit_publish.py \
--video INPUT.mp4 \
--subtitle-file REVIEWED.zh-CN.srt \
--subtitle-language zh \
--prompt-file STORY_CONTEXT.md \
--publish-category lalachan \
--languages fr,zh-Hant,ja,en \
--portrait-blur-fill \
--portrait-blur-mode lalachan \
--logo \
--logo-position top-right \
--platforms shipinhao,youtube,instagram,douyin \
--guided-monitor \
--wait
--subtitle-file imports the cues as both original and polished subtitles and skips Whisper plus implicit AI correction. This also applies when processing is deferred to the serial publish worker; only an explicit --steps value containing transcribe may opt back into Whisper. The prompt remains available as metadata context. Add --correct-subtitles only when the reviewed SRT should still be revised. The import rejects malformed, overlapping, empty, or out-of-range cues.
For unintelligible generated speech, use a short contextual subtitle only after checking the matching visual action. Keep silent scenes silent; do not phoneticize gibberish.
Common Commands
Process and publish in one command with shared reviewed context:
python scripts/lazyedit_publish.py \
--video-id VIDEO_ID \
--use-current-settings \
--prompt-file $LALACHAN_ROOT/references/prompts/PROMPT.md \
--no-correct-subtitles \
--steps keyframes,caption,transcribe,polish,translate,burn,metadata_zh,metadata_en,cover \
--platforms shipinhao,youtube,instagram \
--guided-monitor \
--remote-log-command "ssh $AUTOPUBLISH_SSH 'tmux capture-pane -pt autopub:0 -S -140 | tail -n 140'" \
--wait \
--poll-seconds 10 \
--process-timeout 3600 \
--publish-timeout 7200
Use this for existing videos with no subtitles yet when the user provides background context. The full script is passed as polish_notes; the short metadata brief is passed as metadata notes. Do not reuse the full script as the metadata file. Delete temporary metadata briefs after the run unless the user asks to preserve them.
Publish an already finished output:
python scripts/lazyedit_publish.py \
--video-id VIDEO_ID \
--use-current-settings \
--platforms shipinhao,youtube,instagram \
--no-process \
--wait \
--poll-seconds 10
Publish only YouTube and Instagram:
python scripts/lazyedit_publish.py --video-id VIDEO_ID --use-current-settings --platforms youtube,instagram --no-process --wait --poll-seconds 10
Publish only Shipinhao:
python scripts/lazyedit_publish.py --video-id VIDEO_ID --use-current-settings --platforms shipinhao --no-process --wait --poll-seconds 10
Process then publish:
python scripts/lazyedit_publish.py --video-id VIDEO_ID --use-current-settings --platforms youtube,instagram --wait --poll-seconds 10
Process/publish with lightweight guided monitoring:
python scripts/lazyedit_publish.py \
--video-id VIDEO_ID \
--use-current-settings \
--platforms youtube,instagram \
--guided-monitor \
--remote-log-command "ssh $AUTOPUBLISH_SSH 'tmux capture-pane -pt autopub:0 -S -80 | tail -n 80'" \
--wait \
--poll-seconds 10
Use --guided-monitor when the user wants less manual supervision. It prints heartbeat progress during blocking subtitle correction, follows the local LazyEdit queue, checks the remote AutoPublish queue, and can periodically tail the Pi autopub tmux log. It should not restart services by itself; diagnose first, then intervene only when the queue reports failure or the logs show a clear stall.
If a processed output includes subtitles, confirm the final video path or subtitle-burn config has the logo overlay applied. Normal logo outputs end in _subtitles_logo.mp4.
Override languages for one run without changing Studio defaults:
python scripts/lazyedit_publish.py --video-id VIDEO_ID --use-current-settings --languages zh-Hant,ja,en --platforms youtube,instagram --wait
Create a no-subtitle video while keeping the configured Studio logo:
python scripts/lazyedit_publish.py --video-id VIDEO_ID --use-current-settings --no-burn-subtitles --platforms youtube,instagram --wait
Pure music/audio packaging should go through LazyEdit first, mirroring the
video ZIP contract. LazyEdit creates the metadata, lyrics, audio copy, YouTube
art-track MP4, manifest, original-proof ZIP, and cover candidates; AutoPublish
only consumes the ZIP with publish_shipinhao_music=true and/or
publish_youtube_music=true.
For Musia songs, --lyrics-json is a publication-critical field. Point it to
the final corrected website lyric JSON for the selected audio asset, for
example:
/home/lachlan/ProjectsLFS/Musia/website/data/songs/<media-id>/lyrics/<vocal-set>/<active-lang>.json
Do not pass lyrics.txt, prompt drafts, planning translations, or a companion
language's JSON unless that exact file was corrected against the selected vocal.
If the music package lyrics disagree with fun.lazying.art, rebuild the package
before publishing. Shipinhao Music should display the same corrected lyrics as
the website for that vocal.
Shipinhao Music currently wants plain lyric lines. Keep the default
--lyrics-format plain for Shipinhao packages. --lyrics-format lrc remains a
debug/reference option, but a 2026-06-30 test showed the live desktop form
accepts [00:00.00]... text into the textarea and then rejects the submission
as incomplete. If a proof ZIP stays visible as 0% 删除, it is not ready; wait
for completion, or remove the optional proof and submit without it.
python scripts/lazyedit_music_package.py \
--audio /path/to/song.mp3 \
--title "Song Title" \
--author "Musia 慕莎" \
--language 中文 \
--genre Pop \
--story "Short music story for 音乐人说." \
--lyrics-json /path/to/musia/lyrics/mixed-vocal/mul.json \
--cover /path/to/artwork.png \
--cover-video /path/to/related-video.mp4 \
--cover-count 9 \
--cover-model aginti+codex \
--aginti-cover-count 5 \
--codex-cover-count 4 \
--proof /path/to/website/manifest.json \
--source-url "https://fun.lazying.art/#song-id" \
--output-slug song-title-music \
--platforms shipinhao_music,youtube_music \
--autopublish-url http://lazyingart:8081
The equivalent API is:
curl -fsS http://127.0.0.1:18787/api/music/package \
-H 'Content-Type: application/json' \
-d '{
"audio": "/path/to/song.mp3",
"title": "Song Title",
"author": "Musia 慕莎",
"language": "中文",
"lyrics_json": "/path/to/lyrics.json",
"cover": "/path/to/artwork.png",
"cover_video": "/path/to/related-video.mp4",
"cover_count": 9,
"cover_model": "aginti+codex",
"aginti_cover_count": 5,
"codex_cover_count": 4,
"source_url": "https://fun.lazying.art/#song-id",
"slug": "song-title-music",
"platforms": {"shipinhao_music": true, "youtube_music": true}
}'
Set --post or JSON "post": true only after inspecting the package. As of
2026-06-29, the verified desktop creation route is:
https://channels.weixin.qq.com/platform/post/createMusic
The management/sidebar route is:
https://channels.weixin.qq.com/platform/post/music
Shipinhao currently has no verified standalone desktop 发表专辑 route. Album
(专辑) is handled as required metadata inside the 发表音乐 song form, and
also appears as a read-only management tab. Use AutoPublish
pub_shipinhao_music.py for song creation and pub_shipinhao_zhuanji.py for
read-only album/music tab inspection:
ssh lachlan@lazyingart 'cd ~/Projects/autopub && /home/lachlan/venvs/autopub/bin/python pub_shipinhao_zhuanji.py'
The zhuanji helper saves logs/shipinhao-zhuanji-management.json. Do not
represent a click on 发表音乐 as final listing proof unless the management tab
or backend status shows a row.
After the account has at least one album, the music form may show only
专辑信息 / 选择专辑 / 请选择专辑. The publisher must switch the Vue album
component to 新建专辑 before filling 专辑名称, cover, and intro; otherwise
发表音乐 stays disabled. Verified on 2026-06-29: management showed two albums
and two music rows after publishing One Sky, Three Lights and
アヤちゃん 光の雨.
Do not claim 音乐人说, 歌曲简介, 歌曲故事, or a video-style 声明原创
was filled unless the live create form exposes those fields. In the verified
2026-06-29 run, the story text existed in package metadata and album intro, but
the separate 音乐人说 field was not present; 作品类型 stayed on the page
default 原创, and original proof was uploaded through 证明文件.
Shipinhao music rejects MP3 files below 256kbps. LazyEdit now transcodes low
bitrate MP3 inputs to a package-local *_shipinhao_320k.mp3 copy. Verify with
ffprobe if a package fails to enable the submit button. Required fields filled
by AutoPublish include title, lyrics, author, singer, lyricist, composer,
producer, album name, album description, album cover, original-proof ZIP, and
the 我已阅读《视频号音乐人发表须知》 checkbox.
Shipinhao's 歌曲曲风 dropdown uses Chinese site labels. AutoPublish maps
common English genres such as Bedroom Pop, lofi, and pop to 流行; if a
new English genre times out, add it to pub_shipinhao_music.py instead of
editing the package by hand.
YouTube Music in this workflow means public YouTube Studio upload of a generated
music art-track video. LazyEdit writes youtube_music_video_filename and
video_filename to metadata, and AutoPublish pub_y2b_music.py uploads that
H.264/AAC MP4 with the cover thumbnail, title, description/story, lyrics, tags,
and Musia playlist when available. Direct YouTube Music audio upload is a
personal library feature, so do not claim it as public YouTube Music
distribution.
Music package records are durable in LazyEdit:
python scripts/lazyedit_music_records.py list --limit 20
python scripts/lazyedit_music_records.py update ID --shipinhao-item-url URL
python scripts/lazyedit_music_records.py update ID --deleted
When cover art is not fully prepared, pass one curated cover plus
--cover-video and --cover-count 9; LazyEdit will extract enough frame covers
to fill the nine-background-image package. If AgInTi generates better covers,
pass those as repeated --cover arguments instead.
LALACHAN / AI-Generated Video
Default post-generation rule: when a LALACHAN/Xiaoyunque video has finished,
auto-download it, verify duration/size with ffprobe, copy it to
$LALACHAN_ROOT/Videos, and send the verified MP4 back to the requesting chat.
Submit to LazyEdit only when the current user request explicitly asks for
LazyEdit/import/process or public publishing. For a LazyEdit-only request, use
direct CLI upload with --no-publish; Nutstore AutoPublish import is an
acceptable fallback. For a public publish request, run the normal LazyEdit
publish workflow exactly once for the requested platforms.
If a generated video should go through the normal import path, copy it to Nutstore with a stable _COMPLETED name:
cp -f $LALACHAN_ROOT/Videos/VIDEO.mp4 \
"$NUTSTORE_AUTOPUBLISH/VIDEO_COMPLETED.mp4"
Then watch AutoPubMonitor and find the imported LazyEdit video id:
tmux capture-pane -pt autopub-monitor:0.1 -S -100 | tail -n 100
tmux capture-pane -pt autopub-monitor:0.2 -S -100 | tail -n 100
curl -fsS $LAZYEDIT_API/api/videos | jq '.videos[:20] | map({id,title,created_at,file_path})'
For direct upload with story context:
python scripts/lazyedit_publish.py \
--video $LALACHAN_ROOT/Videos/VIDEO.mp4 \
--title TITLE_COMPLETED \
--use-current-settings \
--prompt-file $LALACHAN_ROOT/references/prompts/PROMPT.md \
--correct-subtitles \
--correction-source polished \
--platforms shipinhao,youtube,instagram \
--wait \
--poll-seconds 10
Use the LALACHAN story/prompt/script as subtitle-correction background. For subtitle correction, treat the script as a reference, not a verbatim source. Use a human middle path: do not over-edit, and do not stay too conservative when the ASR is obviously abnormal, broken, or mismatched with the context. Read neighboring lines, check whether the sentence makes sense, compare it with the audio/Whisper text and the story context, then infer the most likely intended wording. Fix recognition errors, names, objects, and broken phrases while preserving timing and line structure where possible. The final corrected subtitles do not need to be identical to the script if the audio or generated video differs, and they should not invent unsupported content.
For ordinary correction, "preserve timing and line structure" is a hard contract: same cue count, same cue order, and the same start/end timestamps. Changing text is allowed; moving a later sentence into an earlier cue is not. Recovering genuinely missed audible speech requires a fresh audio-alignment pass with evidence-based timestamps, not prompt-driven insertion.
For normal publication, make the story context accurate and readable, then pass
it once with --prompt-file. LazyEdit decides how to use it for subtitle repair
and concise public metadata. Do not pre-compose the metadata or replace the
generated metadata just because a manual version seems preferable.
If an inspected run exposes a real context-separation defect, use temporary
--correction-prompt-file and --metadata-prompt-file inputs as a documented
recovery. Do not commit temporary wrappers, generated ZIPs, or runtime media.
If the user requests no rerun, use --no-process.
Manual Subtitle Quality Pass
After transcription/polish, inspect the polished subtitles before publish when the user gave precise context:
python /path/to/skill/scripts/validate_subtitle_timeline.py \
DATA/VIDEO_FOLDER/VIDEO_mixed.srt \
DATA/VIDEO_FOLDER/VIDEO_mixed_polished.srt
sed -n '1,180p' DATA/VIDEO_FOLDER/*_mixed_polished.md
rg -n "bad term|broken term|ASR artifact" DATA/VIDEO_FOLDER/*_mixed_polished.*
Use a middle path:
- Fix clear recognition errors, broken filler words, wrong objects, and wrong names.
- Keep cue count, cue order, and timestamps unchanged during ordinary text correction.
- Do not invent lines that are unsupported by the transcript/context.
- If the validator reports cue-count or timestamp drift, block publishing. Repair the polished SRT cue-by-cue or fall back to the source ASR when its wording is acceptable. Do not translate or burn a known-misaligned polished file.
- Compare at least one audible line near the start, middle, and end against the rendered subtitle. Structural validation cannot prove audio alignment alone.
- If the corrected text is Chinese filler such as
嗯or呃..., make sure the JSON item language iszh, not a stalejaoren.
When hand-editing subtitles, keep .json, .srt, and .md aligned. Prefer the LazyEdit subtitle-correction save endpoint if it responds quickly. If it starts a duplicate transcription or times out, stop that duplicate process and use the DB recovery note below.
Recovery Notes
A late duplicate Whisper run can write a newer failed mixed transcription row after a valid completed transcript already exists. Symptoms:
process-statusshowstranscribe:error.- Corrected
*_mixed_polished.json/.srt/.mdfiles exist and downstream steps may already be done. - A stray process like
vad_lang_subtitle.py ... --forceorHandBrakeCLI ... _compatible.MOVis visible inps.
First stop only the duplicate worker if it is clearly redundant:
ps -eo pid,ppid,cmd | rg 'vad_lang_subtitle|HandBrakeCLI|scripts/lazyedit_publish.py'
kill PID
Then insert a fresh completed mixed row pointing at the verified corrected files so status ordering recovers:
source .env 2>/dev/null || true
psql "${LAZYEDIT_DATABASE_URL:-${DATABASE_URL:-dbname=lazyedit_db}}" -v ON_ERROR_STOP=1 -c "
INSERT INTO transcriptions (
video_id, language_code, status,
output_json_path, output_srt_path, output_md_path,
error, publication_session_id
) VALUES (
VIDEO_ID, 'mixed', 'completed',
'/abs/path/to/VIDEO_compatible_mixed_polished.json',
'/abs/path/to/VIDEO_compatible_mixed_polished.srt',
'/abs/path/to/VIDEO_compatible_mixed_polished.md',
NULL, NULL
);"
This is a narrow status repair, not a content rewrite. Confirm with:
psql "${LAZYEDIT_DATABASE_URL:-${DATABASE_URL:-dbname=lazyedit_db}}" -P pager=off -c \
"SELECT id, language_code, status, output_json_path, error, created_at
FROM transcriptions WHERE video_id=VIDEO_ID ORDER BY id DESC LIMIT 8;"
If all downstream outputs exist except cover, extract cover directly:
curl -m 180 -fsS -H 'Content-Type: application/json' \
-d '{"lang":"zh"}' \
$LAZYEDIT_API/api/videos/VIDEO_ID/cover | jq .
Then publish without rerunning processing or correction:
python scripts/lazyedit_publish.py \
--video-id VIDEO_ID \
--use-current-settings \
--no-correct-subtitles \
--no-process \
--platforms shipinhao,youtube,instagram \
--guided-monitor \
--remote-log-command "ssh $AUTOPUBLISH_SSH 'tmux capture-pane -pt autopub:0 -S -140 | tail -n 140'" \
--wait \
--poll-seconds 10
Local Section Clip Publishing
Use this path when another workflow has already created a finished section clip, such as a transcript-derived split from Audio2Text. Publish the section clip directly; do not re-split or regenerate the source video.
Before publishing, gather:
- Absolute clip path.
- Suggested title from the section JSON or manifest.
- Section time range.
- Full transcript path, if subtitle correction or metadata needs context.
- Split manifest path, if the user asks what the clip came from.
Example direct publish for a local section clip:
python scripts/lazyedit_publish.py \
--video /path/to/video.mp4 \
--title "Detection, decision making, and automated data generation" \
--use-current-settings \
--platforms shipinhao,youtube,instagram \
--wait \
--poll-seconds 10
If subtitles or metadata should use transcript context, create one reviewed
context note from the transcript and section notes and pass it with
--prompt-file. Split context only to diagnose a demonstrated failure.
For section clips generated from a private source, public metadata should describe only the section content. Do not mention the source recording, private processing notes, or agent workflow unless the user explicitly asks.
Monitoring
Local LazyEdit queue:
curl -fsS $LAZYEDIT_API/api/autopublish/queue | jq '.jobs[:8]'
Remote AutoPublish queue:
curl -fsS $AUTOPUBLISH_QUEUE_URL | jq '.jobs[:8]'
Remote browser automation:
ssh $AUTOPUBLISH_SSH 'tmux capture-pane -pt autopub:0 -S -120 | tail -n 120'
AutoPubMonitor import/session:
tmux capture-pane -pt autopub-monitor:0.0 -S -120 | tail -n 120
tmux capture-pane -pt autopub-monitor:0.1 -S -120 | tail -n 120
tmux capture-pane -pt autopub-monitor:0.2 -S -120 | tail -n 120
tmux capture-pane -pt autopub-monitor:0.3 -S -120 | tail -n 120
Shipinhao Notes
- Shipinhao may require a WeChat QR/login email. Keep monitoring after the user scans.
- Keep the long login wait behavior. It is intentional and useful when the user is away or misses the first QR.
- The automation should wait for upload completion, cover readiness, save draft, then publish.
- Current UI may not expose short title or cover upload; skip those if absent.
- Expected successful log includes
Successfully published on ShiPinHao.
AutoPubMonitor Notes
- Nutstore files copied into
$NUTSTORE_AUTOPUBLISHare synced/imported by AutoPubMonitor. - If a file is renamed while monitor is active, check the tmux panes and queue file before assuming it imported.
- If LazyEdit is down, AutoPubMonitor wrapper must preserve nonzero exit codes so queued files are not silently dropped.
Handoff Checks
Before final response, verify:
curl -fsS $LAZYEDIT_API/api/autopublish/queue | jq '.jobs[:8] | map({id,video_id,status,platforms,remote_status,remote_job_id,error})'
curl -fsS $AUTOPUBLISH_QUEUE_URL | jq '.jobs[:8] | map({id,status,platforms,filename,error,updated_at})'
Report the LazyEdit job id, remote job id, platforms, status, and whether processing was reused or rerun.
Verified Run: 2026-06-03 Typhoon Ping Pong Shark
Request:
- Publish
typhoon_pingpong_shark_duanpian_4x3_15s_2026_06_03_22_46_26_COMPLETEDas before. - Use generated-video subtitle correction from the LALACHAN prompt/script.
Method:
python scripts/lazyedit_publish.py \
--video-id 348 \
--use-current-settings \
--prompt-file $LALACHAN_ROOT/references/prompts/2026-06-03-typhoon-pingpong-shark-duanpian-15s-4x3-budget200.md \
--correct-subtitles \
--correction-source polished \
--platforms shipinhao,youtube,instagram \
--wait \
--poll-seconds 10
What happened:
- LazyEdit had already imported the Nutstore file as
video_id=348. - AI subtitle correction saved polished subtitles before processing.
- Processing completed: transcribe, translate, burn, metadata, cover.
- Shipinhao blocked on login until the user scanned the emailed QR.
- The QR expired once; the existing long wait refreshed it and sent a new email.
- After login, Shipinhao published, then Instagram published, then YouTube published.
Tools used:
scripts/lazyedit_publish.pyfor API/CLI orchestration.curl+jqfor LazyEdit and remote AutoPublish queue checks.ssh $AUTOPUBLISH_SSHandtmux capture-pane -pt autopub:0for remote browser automation logs.tmux capture-pane -pt autopub-monitor:*for Nutstore/import checks.rgon the Pi after installingripgrepfor faster code/log searches.
Final result:
- LazyEdit job
148 - Remote job
job-1780500057985-7 - Platforms
shipinhao,youtube,instagram - Status
done
Verified Run: 2026-06-06 Firefly Cave
Request:
- Copy
$LALACHAN_ROOT/Videos/firefly_cave_cicada_rain_4x3_15s.mp4to Nutstore. - Use
$LALACHAN_ROOT/references/prompts/2026-06-06-firefly-cave-cicada-rain-duanpian-15s.mdas subtitle/metadata context. - Publish to
shipinhao,youtube, andinstagram.
Method:
cp -f $LALACHAN_ROOT/Videos/firefly_cave_cicada_rain_4x3_15s.mp4 \
"$NUTSTORE_AUTOPUBLISH/firefly_cave_cicada_rain_4x3_15s_COMPLETED.mp4"
python scripts/lazyedit_publish.py \
--video-id 352 \
--use-current-settings \
--prompt-file temp/firefly_cave_publish_context.md \
--correct-subtitles \
--correction-source polished \
--platforms shipinhao,youtube,instagram \
--guided-monitor \
--remote-log-command "ssh $AUTOPUBLISH_SSH 'tmux capture-pane -pt autopub:0 -S -140 | tail -n 140'" \
--wait \
--poll-seconds 10 \
--process-timeout 3600 \
--publish-timeout 7200
What happened:
- AutoPubMonitor imported the Nutstore file as
video_id=352. - AI correction saved polished subtitles first.
- Processing completed translation, burn, metadata, and cover.
- Shipinhao required QR login, recovered after login, saved draft, waited for cover readiness, and published.
- Instagram published after crop/Next; YouTube uploaded, filled title/description from metadata, and published.
Final result:
- LazyEdit job
154 - Remote job
job-1780723994544-13 - Platforms
shipinhao,youtube,instagram - Status
done