Social Media Marketing

Social Media Marketing turns one downloadable video into a locally transcribed source and a set of model-generated social drafts. You choose language, platform prompt profiles, tones, model routes, and optional visuals; the Actor records each completed post task as a nine-field Dataset row.

• Download one non-empty video URL through the currently installed extractor.
• Request from 12 languages, 12 platform prompt profiles, 12 tones, and six configured model routes.
• Multiply paid-plan work as platforms × tones × models; duplicate selections repeat work and token events.
• Receive best-effort drafts and up to six shared images, with no publishing, accuracy, compatibility, or engagement guarantee.

Open Social Media Marketing, integrate through the Apify API, or use MCP. Review every generated claim and image before use. The submitter remains responsible for source rights, platform rules, disclosures, and authorized processing.

What does Social Media Marketing do?

The workflow begins by asking yt-dlp to fetch a source video, preferring a merged MP4 no higher than the configured 720p path and trying a direct route followed by public proxies. FFmpeg then extracts audio where possible. Runtime loads the faster-whisper small model on CPU with int8 computation and converts recognized speech into timestamped internal segments.

Those segments are not returned as a transcript. Instead, each paid-plan platform, tone, and model combination becomes a prompt sent through OpenRouter. A task response is appended even when the provider returns zero tokens or empty content. Text-token charging is requested after each call. When the charge manager reports the limit, later tasks stop.

Download recovery tries the direct route and available public proxies with extractor retries, but no user cookie or login field is exposed. Transcription attempts extracted WAV and the original video, up to three times per form. Long media can therefore spend meaningful time in download and CPU recognition before post generation begins, while restricted sources can still fail completely.

Optional image processing happens after post drafts exist. The Actor asks a text model for 20 screenshot timestamps, stores available PNG frames, and asks another model call to select six filenames. If selection fails, it randomly chooses up to six stored screenshots. Original returns selected screenshot URLs. Eight styled choices attempt an image-to-image call for each selection. All final rows share the same resulting images list.

How do I run Social Media Marketing?

Use the smallest configuration first: one platform, one tone, one model, no images. This isolates video download, local transcription, and a single post call.

{
  "video_url": "https://www.tiktok.com/@openai/video/7521583654028512525",
  "image_style": "None",
  "language": "English",
  "platforms": ["X-Free"],
  "tones": ["Educational"],
  "models": ["DeepSeek"]
}

The Console prefill does not make fields optional: all six keys are required. Wait for completion, then check status, token events, Dataset rows, and KVS keys. Terminal success alone is insufficient because the main handler catches download, transcription, validation, and provider failures. Token events can also precede a final zero-row Dataset.

Free-plan runtime narrows the request automatically to DeepSeek, the first platform, the first tone, and no images. Paid-plan behavior uses the full task product unless a budget limit or failure interrupts it.

Before enlarging the request, remove accidental duplicates in the three arrays. The schema has no maximum array length, and runtime does not deduplicate, so repeated valid labels can create an unexpectedly large Cartesian product. Ten copies of one platform, two tones, and two models request 40 calls even though only four distinct combinations exist.

What data does Social Media Marketing return?

Each saved task row has exactly nine fields: Actor provenance, creation time, post-call token count, model label, platform label, requested language, requested tone, a shared image URL array, and generated content.

{
  "processor": "https://apify.com/truefetch/social-media-marketing?fpr=aiagentapi",
  "processed_at": "2026-07-23T20:30:00+00:00",
  "tokens": 816,
  "model": "DEEPSEEK",
  "platform": "LinkedIn",
  "language": "English",
  "tone": "Professional",
  "images": [],
  "content": "A draft generated from the locally recognized video speech."
}

tokens is only the OpenRouter total_tokens reported for that post call. It excludes faster-whisper work and the separate text/image calls used for visual selection and styling. platform describes an internal prompt profile; nothing is posted to that service. content may be empty. Signed image URLs depend on retained KVS records and access conditions.

The input URL and internal transcript are absent from rows. Preserve the run ID and input separately when you need lineage or source comparison.

Image storage is best-effort. None should leave only the platform-managed INPUT record. Original mode retains final selected screenshot records. Styled modes retain successfully uploaded styled images and attempt to delete intermediate PNGs. The returned links are signed KVS URLs, not a promise of permanent public hosting.

What inputs can I use?

video_url is any non-empty string at schema level. Compatibility is decided by the extractor and the live source. Private, deleted, login-gated, geoblocked, unsupported, or changed video pages can fail.

image_style supports None, Original, Memes, Realistic, 3D, Anime, Cartoon, Cyberpunk, Digital Art, and Pixel Art. None skips visual work. Original returns selected source frames. Other values are instructions sent to a configured image model and can produce fewer than six or no images.

language supports English, Chinese, Japanese, Korean, Spanish, French, German, Russian, Portuguese, Italian, Arabic, and Hindi. It is a generation instruction, not a verified language result.

platforms, tones, and models are required non-empty arrays. Platform values are X-Free, X-Premium, Instagram, LinkedIn, Facebook, Pinterest, Reddit, TikTok, YouTube, Rednote, Threads, and Mastodon. The 12 tones range from Questioning and Emotional to Educational, Professional, and Casual. Models are DeepSeek, Claude, Gemini, Grok, Qwen, and GPT. Runtime does not deduplicate arrays.

What platforms and markets are supported?

The 12 output platform names are writing profiles only. They do not authenticate to, query, schedule, or publish on those networks. Internal prompt character targets can drift from platform rules, subscription limits, content types, or regional policies, so verify the current destination requirements independently.

Input-video coverage follows the current yt-dlp build. Its official supported-sites document warns that even listed extractors are not guarantees. Output markets are represented by the 12 language prompts; generated localization, cultural fit, and regulatory suitability require qualified review.

Why use Social Media Marketing?

The Actor centralizes download, transcription, prompt expansion, model calls, image extraction, storage, billing events, and Dataset output. It can help teams compare controlled draft variants without building those orchestration pieces separately.

It also exposes meaningful audit points: the selected labels, provider-reported token count per post, final images, run logs, charged events, and storage identifiers. The documentation distinguishes these observations from claims the runtime cannot prove, such as factual accuracy, engagement improvement, platform acceptance, or rights clearance.

Who is Social Media Marketing for?

Editorial teams can produce candidates for a review queue. Localization teams can request an initial target-language draft and then verify terminology, meaning, names, dates, and cultural context. Marketing operations teams can compare tones or model routes while retaining the exact source run.

Rights holders can select Original frames or request visual transformations for authorized media. The Actor is not suitable for impersonation, fabricated endorsements, deceptive political content, harassment, or unauthorized reuse of copyrighted, private, or sensitive recordings.

A practical review should cover retrieval authority, transcript fidelity, factual claims, language quality, tone, disclosure, image provenance, and the current destination-platform rules. Keep the original source available to the reviewer and record which row was approved. Terminal success, a token count, or a visually polished image is not an editorial or compliance decision.

Can I use Social Media Marketing through an API or MCP?

Yes. Use Actor ID ay04FddyeCI9hkXJG or truefetch/social-media-marketing. Keep the Apify token outside source code and calculate task multiplication before execution.

{
  "method": "POST",
  "url": "https://api.apify.com/v2/acts/truefetch~social-media-marketing/runs?waitForFinish=300&maxTotalChargeUsd=1",
  "headers": {
    "Authorization": "Bearer YOUR_APIFY_TOKEN",
    "Content-Type": "application/json"
  },
  "body": {
    "video_url": "https://www.tiktok.com/@openai/video/7521583654028512525",
    "image_style": "None",
    "language": "English",
    "platforms": ["LinkedIn"],
    "tones": ["Professional"],
    "models": ["DeepSeek"]
  }
}

Read items from defaultDatasetId. Some MCP plans cannot directly execute every public Actor; authenticated REST can start an exact build, while MCP can independently read back Actor details, run state, Dataset items, and KVS records.

How much does Social Media Marketing cost?

The current TrueFetch schedule lists Actor Start and Actor Usage at $0.0001 per event, Text Token at $0.000034, and Image Token at $0.000080. Text tokens include post calls and, when visuals are enabled, timestamp and screenshot-selection calls. Image tokens apply to styled-image calls; Original screenshots do not use that event.

One 1,000-token post adds $0.034 in text-token events. Twelve post calls reporting 12,000 total tokens add $0.408, before usage, start, and visual work. An image-enabled run can cost more even if final image storage fails. The configuration requires a minimum maximum-charge value of $1.

Active pricing at execution is authoritative. A charge cap stops further charge acceptance; it cannot promise a complete task grid or final Dataset rows.

How does Social Media Marketing compare with alternatives?

A transcription Actor returns recognized speech; this Actor keeps its local transcript internal and returns social drafts. A caption downloader retrieves source-exposed subtitle files. A media downloader returns the asset. A scheduler publishes approved content. A design tool offers manual visual control.

Choose this Actor for draft orchestration from one authorized video. Use specialized tools when you need a transcript as the deliverable, lossless media, manual design precision, platform authentication, scheduling, or publication approval.

What are the limits and troubleshooting steps?

If no rows appear, validate the URL, source accessibility, download logs, FFmpeg output, and transcription stage before investigating OpenRouter. If a row has empty content, the task handler likely caught a provider error. Model identifiers are pinned in runtime and can disappear or change behavior.

For images, check whether the account is paid, style is not None, screenshot files were created, timestamp JSON parsed, KVS uploads succeeded, and the selection/styling calls returned usable data. Cleanup retains INPUT and final image filenames best-effort while deleting intermediary records.

Use current OpenRouter API documentation, faster-whisper project documentation, and yt-dlp documentation for their respective layers.

FAQ

Does it publish posts? No. It only creates Dataset drafts and optional image records.

Are the model choices always available? No. They map to runtime-pinned OpenRouter IDs and can fail or change.

Why are the same images on every row? Visual selection runs once after post generation and the final list is copied into every post object.

Can arrays contain duplicates? Yes. Runtime does not remove them, so paid-plan tasks and charges repeat.

Is the local transcript returned? No. It is used internally as prompt context.

Can a charged run return zero rows? Yes. Charges occur before the final Dataset batch.

Are six images guaranteed? No. The result can contain zero to six final URLs.

Related TrueFetch Actors

• Video To Text — choose a workflow where recognized text is the primary output.
• Instagram Video Transcript — target authorized Instagram transcript use cases.
• TikTok Video Downloader — retrieve authorized TikTok media instead of generated social drafts.

Support and last updated

For support, provide the run ID, UTC time, build, redacted URL, input selections, expected task count, terminal state, event counts, Dataset count, KVS keys, and the failed stage. Do not share Apify or OpenRouter keys, private media, or unredacted sensitive output.

Contact the TrueFetch group or support channel. Last updated: July 23, 2026.