The content creation agent: research, script, video, captions, five platforms

Use this stack as a page: the omnichannel content agent.

The “content creation agent” card sells a fully automated path: ideation → creation → publishing, then repeat. The dashboard always has a Run Agent button. That button is an orchestrator. It is not a new kind of intelligence.

If you already read faceless shorts, this is the wider loop: more research up front, captions as a named step, and five outlets (TikTok, Instagram, YouTube, LinkedIn, X) instead of two. Same warning: auto-post is a timer, not a strategy.

The stack:

1. Topic research — Perplexity, optionally NotebookLM if the ideas must stay inside your sources.
2. Script — Claude or ChatGPT. One writer.
3. Video — Kling AI, Sora, Runway, Pika, Veo 3, or Fliki. One camera.
4. Voice — ElevenLabs unless the video model’s native audio is the brand.
5. Captions — burned in, or a subtitle file. Not a second script by a second model.
6. Publish — n8n / Zapier to the five networks. Those networks are not /tools pages.

This is not our agent product. It is the job list so “fully automated” does not mean “one ChatGPT tab forever.”

Why a single Run Agent is a lie unless the jobs are split

A model that picks a topic, writes a script, generates a clip, captions it, and posts it will optimize for finishing the Zap, not for being right. The research step will invent a trend. The video step will ignore the script. The LinkedIn cut will be the TikTok file with a worse crop.

Perplexity exists so research has sources. Pair it with the rest of the research habit in the research stack if the agent is supposed to be factual. NotebookLM is the “only these PDFs” box — use it when the channel is a product you actually ship, not a vibes newsletter.

Claude ($20/month paid, 1M+ context) should see last week’s posted scripts so it does not repeat the same hook. That is the whole point of a long thread. ChatGPT ($20/month paid) is the swap, not a second writer in parallel.

The six jobs (and what each must not do)

Research owns novelty, not the voiceover

Perplexity: “what is getting search and social attention in this niche in the last 7 days,” with links. NotebookLM: “what did we claim in the white paper.” Merge in the orchestrator, not in a vibes prompt.

Do not let the video model pick the topic. Do not scrape a competitor’s entire channel into a public model if you are not allowed to.

Script owns words for this slot

Output three lengths from one brief: 30s (TikTok/IG), 45–60s (Shorts/Reels), 90s (LinkedIn/X may still want the short one). Same thesis. Different density.

Do not generate five unrelated scripts and call it a calendar.

Video owns the file

Kling AI: free credits, paid from $10/month. Sora: gated, no monthly number on our matrix. Veo 3: Flow free pool, Google AI Pro $19.99/month. Runway $12/month paid after free. Pika $10/month. Fliki $10/month.

YouTube may want 16:9. TikTok wants 9:16. The agent that “publishes everywhere” without a crop step is how you ship letterboxing.

ElevenLabs owns the voice you will still want in December

Free row, paid from $5/month. Native Kling/Veo audio is a demo. Brand voice is a library.

Captions own accessibility and mute-autoplay

Generate from the final VO, not from the script draft you rewrote in step 2b. If you use YouTube auto-captions, check them. Wrong captions on a “fully automated” agent is a product bug, not a platform bug.

n8n owns the clock

n8n ~$22/month cloud. Zapier $19.99/month Professional. Schedule is a table: platform, time, aspect ratio, caption text (the social caption, not the burnt-in subtitle).

LinkedIn and X are not TikTok. A faceless listicle can work on all five. A kitchen UGC might not. The agent should have a skip flag per platform, not a blind fan-out.

One cycle, eight steps

1. Perplexity (and NotebookLM if needed) produce a topic packet with URLs.
2. Human accepts or kills the topic. This is the step people delete, then regret.
3. Claude writes the three lengths + social captions.
4. One video model renders the picture. ElevenLabs renders VO.
5. Captions from the VO.
6. Export 9:16 and 16:9 if YouTube is in the set.
7. n8n publishes. Log URLs back to a sheet.
8. Next cycle starts from that sheet so Claude can see what already ran.

“Repeat automatically” without step 2 is how you get banned or boring. Keep the kill switch.

What this stack is not

It is not the inbound voice agent. Different loop.

It is not five affiliate CTAs in one caption. If you mention a tool, link it like we do here or do not mention it.

The page with the four matrix slots is the omnichannel content agent — presenter in HeyGen, not the faceless pipe.