had a random thought this morning:
what if i could automate a decent chunk of a short-form content pipeline using n8n, openrouter, supabase, and a handful of APIs stitched together.
spent most of the day building it, currently have workflows running that:
> ingest story candidates from external sources
> enrich + validate them through llm-driven research pipelines
> generate structured research packets with normalized outputs
> transform research into short-form script artifacts
> persist intermediate stages into supabase for downstream processing
> handle structured json outputs, parsing, and workflow state transitions
- all orchestrated through n8n.
the original plan was to take it all the way through automated video generation as well.
spent a good chunk of time experimenting with heygen and a few adjacent tools for the final stage, assuming it would mostly be an orchestration problem. turns out the orchestration is the easy part.
the harder problem is that video quality is still heavily dependent on editing decisions.
you can generate an avatar, voiceover, and even rough scene sequences automatically, but getting something i'd actually publish still requires manually reviewing shot selection, pacing, cut timing, visual continuity, b-roll choices, scene transitions, emphasis points, and overall narrative flow.
the generation layer works.
the editing layer is where most of the quality still comes from.
initially jumped into it thinking the challenge would be wiring services together but ended up realizing the bigger challenge is replacing the hundreds of small editorial decisions that happen between a script and a watchable final video
to sum up, a "small experiment" somehow turned into multiple workflow stages, database schemas, structured-output enforcement, retry mechanisms, state management, prompt iteration, and a surprising amount of debugging around edge cases.
fun rabbit hole though.
now back to the actual work i was supposed to be doing lol.
English
