it-reelsmaker-online
Online sources for it-reelsmaker video edits: vertical stock footage (Pixabay, Pexels, Magnific), CC-licensed memes (Openverse; GIPHY as reference only), paid AI video generation (Veo, Kling, LTX via fal.ai) for B-roll, a cloud transcript (OpenAI), subtitle translation through the Claude, OpenAI or Gemini API, a narration in the person's own cloned voice (ElevenLabs), and figure cut-outs on your own server. Use only inside an it-reelsmaker edit, and only when the person names a cloud service, an API, paid generation, stock or online sources, or their own server, or turned these on: "find stock footage", "generate an insert", "online memes", "transcribe it through OpenAI", "translate the subtitles through Gemini", "voice the script with my cloned voice", "cut out the figure on my server". A plain "transcribe", "translate the subtitles" or "cut out the person" belongs to the core skill, which does it locally. Every download and paid run waits for the person's approval of the source, size and price.
- License 许可
- MIT
- 最近更新
- 2026-11-14
- Stars
- ★ 9
- Forks
- 0
01 前言 Front Matter
- name
- it-reelsmaker-online
- description
- Online sources for it-reelsmaker video edits: vertical stock footage (Pixabay, Pexels, Magnific), CC-licensed memes (Openverse; GIPHY as reference only), paid AI video generation (Veo, Kling, LTX via fal.ai) for B-roll, a cloud transcript (OpenAI), subtitle translation through the Claude, OpenAI or Gemini API, a narration in the person's own cloned voice (ElevenLabs), and figure cut-outs on your own server. Use only inside an it-reelsmaker edit, and only when the person names a cloud service, an API, paid generation, stock or online sources, or their own server, or turned these on: "find stock footage", "generate an insert", "online memes", "transcribe it through OpenAI", "translate the subtitles through Gemini", "voice the script with my cloned voice", "cut out the figure on my server". A plain "transcribe", "translate the subtitles" or "cut out the person" belongs to the core skill, which does it locally. Every download and paid run waits for the person's approval of the source, size and price.
02 资源文件 11
- 参考文档references/generation.md4.3 KB
- 参考文档references/sources.md19.0 KB
- 参考文档references/voice.md6.3 KB
- 脚本scripts/genfootage.py22.2 KB
- 脚本scripts/matte_server.py13.8 KB
- 脚本scripts/memes_online.py9.7 KB
- 脚本scripts/reels_online.py25.4 KB
- 脚本scripts/stock.py6.5 KB
- 脚本scripts/transcribe_online.py30.6 KB
- 脚本scripts/translate_online.py11.3 KB
- 脚本scripts/voice_online.py69.3 KB
03 SKILL.md 正文
Online sources for it-reelsmaker
An add-on to the it-reelsmaker skill. It plugs three online sources into the core's visual plan (step 7a and the core skill's inserts reference): stock footage, online memes and a video model; and a cloud recognizer into the core's step 2 (transcript). It is never used without the core. Everything not described here follows the core's rules: why an insert is needed, the intensity budget, meme size and placement, clip preparation and plan checks.
Connect it to the project
The core's scripts are the only entry point; they load this add-on as a library. Link it to the editing project once, and again after every add-on update (the plugin folder changes with the version; the core then warns that the add-on is not found):
cd {{PROJECT_ROOT}}
python "${CLAUDE_SKILL_DIR}/scripts/reels_online.py" link
It writes online_scripts into the project's it-reelsmaker.json (unlink removes it). After that the core does the rest: footage.py search, plan-search and pick include the stock providers, reelcfg.py show reports what turns on and why not, and the add-on's own commands run through the core runner (<core scripts> is the core skill's scripts/ folder):
| Command | What |
|---|---|
python <core scripts>/addon.py memes search "facepalm" [--provider openverse|giphy] |
online meme candidates |
python <core scripts>/addon.py memes fetch openverse:<id> --yes |
download an approved CC image into memes/online/ with its attribution |
python <core scripts>/addon.py gen manifest|validate edit/<id> |
model prompts (generated/prompts.md, prompts.json) with the price |
python <core scripts>/addon.py gen run edit/<id> --yes |
paid generation, only after the approved price; the clip is picked up by the core's codescene.py ingest |
python <core scripts>/addon.py matte cut edit/<id> --from … --to … [--host H] |
figure cut-out on your own server (below) |
python <core scripts>/addon.py transcribe edit/<id> <source> --provider openai [--language ru] [--yes] |
a more accurate transcript (below); --price only names the price |
python <core scripts>/addon.py translate edit/<id> --to en --provider anthropic|openai|gemini [--yes] |
subtitle translation through an API (below) |
python <core scripts>/addon.py voice list|use|clone|plan|run|status|build|check … |
a narration in the person's own cloned voice (below, references/voice.md) |
Reel settings
Online sources are off until turned on for a video (reelcfg.py save edit/<id> --set use_online_footage=true) or for the whole project. Keys live in edit/<id>/reel.json, with the same layers as the core, each overriding the previous: the core's defaults with this add-on's → the brand tone preset → the project's reel-defaults.json → brand.json → inserts → reel.json → words in the prompt:
{
"use_online_footage": true,
"use_online_memes": false,
"generation_engines": ["code"],
"generate_now": false,
"broll_priority": ["project", "local", "online", "generated"],
"meme_priority": ["local", "online"],
"online_footage_providers": ["pixabay", "pexels"],
"online_meme_providers": ["openverse"],
"generation_provider": "fal",
"generation_model": "veo3.1-fast",
"generation_seconds": 6
}
| Key | Meaning |
|---|---|
use_online_footage |
online stock as a B-roll source |
use_online_memes |
online memes — CC with commercial use only |
generation_engines |
code — code scenes (core), model — a video model. Default ["code"]: installing the add-on does not change the core's code scenes; ["code", "model"] (for the video or the project) writes model prompts too |
generate_now |
whether to start paid generation; without it only a ready prompt is written and the insert stays pending |
online_*_providers, generation_* |
providers, model and clip length (references/sources.md) |
Priority. B-roll: project → library → online stock → generation (code or model) → main footage. Memes: own folder and library → online CC → no meme.
What actually turns on. Before the plan, check availability: an enabled source with no key, no network or in offline mode switches itself off, and the reason goes into the report. A source failure (no key, no network, a 400/401/500 response) is a warning, not a crash: the edit always reaches the render with whatever is available.
Rules
- Only after a “yes”. The plan is shown to the person with the list of downloads (source, MB) and paid runs (≈ $). Downloads and generation happen only after the person approves that plan. A first generation run without approval only names the price.
- Network and keys. Check offline mode and the key before every network action, not only at search time. Keys come from environment variables or from a file outside the project (
references/sources.md). Never ask the person to paste a key into the chat, and never runkeys setyourself: to add a key, the person runspython <add-on scripts>/reels_online.py keys set <NAME>in their own terminal (hidden input; it refuses to run without a terminal); you may runkeys list(values masked) to see what is set. Mask API keys, authorization headers andkey=/token=in URLs in all output and saved errors; keys are never written to project files or git. - Stock. Download only what was selected. Record attribution in
edit/<id>/inserts/credits.jsonand in the post caption. Never present a stock person as “our client or candidate”: stock licenses forbid implied endorsement. - Online memes. Embed only CC0, PDM, CC BY and CC BY-SA, with attribution; check the license on the image's detail page. GIPHY is a reference only (an Instagram sticker or a re-shoot). Film stills, celebrity photos and other people's memes featuring people — never.
memes fetchnames the file by its bytes, not by the link: emoji sets often come as SVG, whichmemes.pycannot read and this add-on cannot draw without native libraries. Search marks SVG results; prefer a raster one. A fetched SVG stays as.svgnext to its sidecar (license, attribution): convert it yourself to a PNG of the same name (≥ 600 px, transparent background) with a tool you trust, thenmemes.py index. - Generation. Prompt fields and template, checks, and “money is never spent twice” —
references/generation.md. - AI label. A photorealistic insert from a model → turn on the “AI info” label when publishing (Meta rules). Stylized code graphics don't need it.
- Brand restrictions (
forbidden_imagery,forbidden_imagery_en) apply to stock queries and generation prompts.
Cloud transcript (core step 2)
The core transcribes on this computer; its word times match the audio, its words are sometimes wrong (a case
ending, a word dropped at a joint between takes, and wrong words become visible in burned-in subtitles). addon.py transcribe keeps the local word times and lays a cloud recognizer's text onto them: the same word gets the cloud's
spelling and punctuation, a different word in its place is replaced (the local one kept in local), a word only the
cloud heard gets a time between its neighbours (est: true), and words only the local model heard stay: often the
repeat of a retake the cloud tidied away, so the report lists them as places to listen to (core step 3). Read the
report's lines before the cut plan. The local transcript is made first if missing and kept as <stem>.local.json.
- Provider:
openai: textgpt-transcribe($0.0045 per minute of audio). OpenAI's text models return no word times; itswhisper-1does, but measured on real footage they were off by more than 0.15 s for one word in five, so they are used only without a local model (--words cloud, $0.006 per minute). Files up to 25 MB, about 13 minutes of the core's 16 kHz WAV; a longer one goes as mono mp3 (32 kbit/s: ~100 minutes, 16 kbit/s: ~3.4 hours, so a whole long recording forlongform.pyfits), packed in a temporary folder and removed.groq: textwhisper-large-v3-turbo($0.04 per hour, a free tier), as clean asgpt-transcribeon the Russian test video; its word times are not used (off by more than 0.15 s for almost half the words, pauses between phrases lost), so it gives the text only, on the local times, and--words cloudis refused for it. Settranscription_fallback=groq(reelcfg.py defaults --set …) and aGROQ_API_KEYto use it when OpenAI fails; or run it directly with--provider groq. - Audio: the core's
edit/<id>/audio16k-<source stem>.wav, on the video's timeline (never the video file). It leaves the computer: say so when offering it. - Paid, so only after a “yes”: name the price (
--price), run with--yes; or the person setstranscription_provider=openaiin the project defaults (reelcfg.py defaults --set …) once, which is their standing choice for every video. - Fallback: no key, no network, no credits, a refused request → a warning, exit code 2, the local transcript
stays as it was; the edit goes on with it. Key:
OPENAI_API_KEY, added by the person withkeys set.
Subtitle translation through an API
By default you translate the phrases yourself in the session (core: subs.py, free, nothing leaves the computer).
addon.py translate is for a translation the person wants from a particular model, or a second opinion: it sends
the phrases (text only) with the brand's voice and forbidden_words, asks for one translation per phrase within the
characters that can be read while it is on screen, in the speaker's form of address, writes them into the empty
phrases of subs/<lang>.json (--force: all) and runs the core check (subs.py apply). Read the printed pairs: a
model may shorten away meaning. Providers and default models: anthropic claude-sonnet-5-5, openai gpt-6.1-sol,
gemini gemini-3.8-flash (--model or the translation_model setting for another). A video is a few thousand
tokens (the command prints them; the price is the model's rate); paid all the same: --yes, or the person's standing translation_provider setting. Keys:
ANTHROPIC_API_KEY, OPENAI_API_KEY, GEMINI_API_KEY (the person adds them with keys set). A phrase the model
skipped stays empty: translate it in the session, then subs.py apply.
A narration in the person's own voice (ElevenLabs)
For a reel with no voice recorded (the core's “no footage” question, or scenes and silent shots) the person may want their own cloned voice to read the script. It is a narration like a voice note from a phone: the core cuts it as an off-camera voice. Offer it only when the add-on is linked and the person names it or has no way to record.
- Only the person's own voice, and only after they confirm it: read them the consent sentence the command prints and
pass
--consent ownafter their explicit “yes”. Someone else's voice (a colleague, a client, a family member) is not supported in this version, even if the voice is in their account. A voice ElevenLabs asks to verify is blocked. - Start with the account's voices:
voice list(free) shows them with a link to listen; the person says whose voice it is; thenvoice use <voice_id> --name <name> --consent own. A new clone (voice clone) only from a fresh, clean 1–3 minute recording of the person alone, which they chose and listened to; it goes to ElevenLabs. - The script is the brief's text, word for word:
voice script edit/<id> --text <file> --voice <name>(blocks by paragraph, ≤ 1000 characters each). Numbers, dates and abbreviations: best written in words. - Paid, the price first:
voice plan edit/<id>names credits per block and a plan id; show it, and only after the person's explicit “yes” to that price runvoice run edit/<id> --yes --plan <id>(generate_voice_now=truefor the video). A changed text, voice or setting is a new plan and a new “yes”. A block leftunknownis never sent again by the script: follow its message (history,recover,clear). - Listen before editing: the person listens to
generated/narration.wavwhole; stress and pronunciation are checked by nothing else (voice checkcompares words only). Then picture blocks or scenes as for a recorded narration (cut.py, thenvisual_plan.py init edit/<id> --scenes-voice). - Label when publishing: realistic synthetic audio may need the platform's AI label (Instagram's “AI info”); say so in the report.
Key ELEVENLABS_API_KEY, added by the person with keys set; it needs the permissions listed in references/voice.md.
Figure cut-out on your own server
The core cuts the figure out on this computer (matte.py; the cut-out reference of the core skill, figure.md). On a weak laptop the same step can run on a server you control: addon.py matte cut … --host ${user_config.matte_host} (the add-on setting). If that value is empty or still reads ${user_config…}, it is not set: leave --host out entirely (an empty --host breaks the command line), and the script takes matte.host from it-reelsmaker.json or REELS_MATTE_HOST; if neither is set, ask the person once and write matte.host there. Frames of the span go to that host as one JPG archive over your own SSH client, the WebM with alpha comes back, and the temporary files are removed there, on error too. The server needs rembg with the u2net_human_seg model (its path on the server: matte.server_rembg, or matte.rembg when you never cut out locally); one cut-out at a time under a server-side lock, with a memory check. Nothing goes anywhere else.
References
| File | When |
|---|---|
references/sources.md |
connecting stock, video models and online memes: terms, prices, endpoints, keys |
references/generation.md |
model generation: prompt fields and template, checks, run rules, AI label |
references/voice.md |
a narration in a cloned voice: files, states, recovery, joining, the contract with the core |
Paid submissions reserve gen.job under the plan lock before contacting fal. A submitting or unknown state never resubmits automatically: check the provider dashboard, then restore the request ID and queue URLs to resume, or clear the job by hand only after confirming that no paid request was accepted. Resumed fal jobs use only FAL_KEY; other saved providers are refused. Downloaded raw clips are bound to their request ID, so an older clip cannot satisfy a new job. Requests and redirects require HTTPS; credentials are stripped on an origin change, and fal queue hosts are checked at every hop. JSON responses are limited to 20 MB and downloads use unique temporary files.
04 同仓库更多 Skills
it-reelsmaker
Edit vertical short videos (Reels, Shorts, TikTok) from talking-head, two-person or horizontal footage, cut a long recording (podcast, webinar, interview) into clips, lay voice notes over your own shots, or make a video with no footage: a brand promo, or designed scenes over a voice note from your phone, optionally with a brand mascot drawn in code. Local transcription, pauses and retakes cut by audio, virtual camera and on-brand motion graphics in Remotion, word-timed subtitles and their translation, optional inserts (B-roll, code scenes, memes), face-aware layout, −14 LUFS mastering, 1080×1920 render. Use when asked to edit, cut or fix a vertical video: “edit a reel”, “make a Reel from IMG_xxxx.MOV”, “clips from this podcast”, “add subtitles / a hook / a cover / B-roll”, “translate the subtitles”, “a promo without footage”, “a video from my voice note”, “add a mascot”, “the video jumps back”. Not for long horizontal edits, format conversion or compression. Needs ffmpeg, Node.js with Remotion and Python.
graders
暂无简介
graders
暂无简介
graders
暂无简介
graders
暂无简介
graders
暂无简介