One content engine, swap a file to change the whole channel.

persona.md  ->  the only file you change per channel

  generate   face-consistent stills + short clips (MCP video gen)
  edit       agent-driven cut, captions, render (local editor)
  post       finished reel  ->  human review gate

build the engine once, run many niches on it

Every niche-specific detail (the character, the voice, the angle, the visual rules) lives in one per-persona file the pipeline reads at run time, so launching a new channel is configuration, not a rewrite. The stages are split by cost and reliability: generation produces a face-consistent presenter from a reference image so the same character appears across every scene, a companion editor that Claude drives through tool calls handles the assembly, and a human gate sits before anything is posted.

A teardown of real creators in the space taught me the counterintuitive part: what the presenter says drives results far more than raw generation quality. So the pipeline treats the script as the primary artifact, transcribing a strong in-niche reference and rewriting it with an LLM pre-loaded with the persona. The value is not any single AI model, it is the plumbing that chains a face-identity model, a video generator, a voice model and a programmatic editor into one flow, and the judgment about where to spend metered credits versus where to draft cheap.

September upgrade: from a pipeline to a workflow. The first version could make a reel. It could not tell me which reel to make next, or make ten of them without drifting. So I added a shared reference library that turns a viral reel into a scene-by-scene rebuild spec, and every reference is classified by measurement rather than by eye: ffmpeg scene detection counts the cuts, and the count decides the generation method, because one scene is one generation and the method sets the cost. A ledger stamps each reference as used in the same edit that writes the post, so a spent idea is never offered twice. The first batch took 24 references to 11 finished reels across two personas in one session.

Most of the hard part was finding where the tools quietly disagreed with the prompt. Attaching a reference photo pulled the source creator's room into the output whatever the text said, so pose references are now background-removed cutouts and each persona's room has its own locked reference image. A generation call once returned "completed" with the character's identity silently missing, because the connector drops one argument shape without an error, so the standing check is to read the parameters the API echoes back, not trust that the call was accepted. And "keep the framing locked" in a prompt did nothing, so the still is cropped before it is animated: pixels constrain the composition, words do not.

• multi-tool AI orchestration
• MCP media generation
• identity-consistent generation
• config-as-persona design
• cost-aware pipelines
• human-in-the-loop
• measurement over guesswork
• batch production

multi-week build

What it actually took. My first instinct was to chase perfect-looking footage, and that was the wrong instinct. The creators actually making this work spend their effort on the script and keep the generation cheap. Rebuilding the pipeline around that order, script first and generation rationed, was the real lesson.

Private build.