Tutorial publish pipeline

Summary: A skill that runs after recording a tutorial video, takes the transcript as input, and produces the full publish package: title options, video description, chapters, tags, thumbnail image with on-image text, a LinkedIn post promoting the video, a YouTube Shorts script derived from a key moment, and a derivative topic note for the newsletter. The user reviews each artefact, selects between options where the AI offered them, and publishes.

What makes this automation-level: One trigger fans out into multiple deliverables that previously required separate prompts and tool-switching. The user reviews each output but doesn't drive each artefact's generation — the orchestrator handles sequencing, format, and brand consistency. Each artefact would be a good AI candidate on its own; the leverage comes from running them together off a single source so they stay coherent with each other.

Map

The recurring workflow is publishing a tutorial. Decomposed into: process the transcript, write title options optimised for searchability, write a long-form description, generate timestamped chapters, suggest tags, produce a thumbnail image with on-image text, write a LinkedIn promotional post, identify a high-leverage moment for a Shorts script, draft the script, write a derivative topic note for the newsletter that points back to the video.

Every artefact is a good AI candidate — they're all transformations of the same source material with format-specific constraints. The decision of which video to publish and final approval of each artefact stayed with the human; the production work got delegated.

Test

Initial test was a manual one-off prompt for each artefact, each in its own session. Quality was acceptable but the artefacts drifted from each other — the title described a slightly different angle than the LinkedIn post, the Shorts script focused on a moment the description didn't highlight.

The fix was orchestration with shared context: a skill that processes the transcript once, extracts the core thesis and the standout moments, then generates each artefact downstream from that shared analysis. Coherence improved immediately because every artefact was working off the same source-of-truth understanding of what the video was about.

Integrate

Built as a skill triggered after a recording is done. Inputs: the transcript and a brand visual identity reference. The skill processes the transcript first, identifying the thesis, the three or four standout moments, and the natural chapter breaks. From that base, each artefact gets generated in sequence with format-specific instructions.

The thumbnail uses the brand visual identity (colours, typography, layout templates) so it stays on-brand without re-specifying every time. The LinkedIn post and Shorts script reference the same moments; the description and chapters share timestamps; the newsletter topic note links back to the video and reuses the thesis as its hook.

Compound

Two compound mechanisms.

The artefact templates refine through performance feedback. Title options that perform well in click-through become templates the AI defaults to; ones that underperform get noted as patterns to avoid. The same applies to LinkedIn post structure, Shorts hook patterns, and thumbnail compositions. Over time the AI's first-draft output trends closer to what consistently performs without the user having to re-specify each round.

The brand visual identity also compounds. Every thumbnail correction the user makes — adjusting positioning, changing emphasis, swapping a typeface — becomes input that refines the brand templates. New thumbnails inherit the corrections automatically because the templates themselves got tighter, not because the AI got smarter.

example-image-generation — The thumbnail generation step is a smaller-scale instance of the iterative-refinement pattern, embedded inside this larger orchestrator
compound-engineering — Performance-feedback refinement of templates is one of the cleanest forms of compound improvement, because the signal (engagement metrics) is direct and observable
agent-design-principles — Generating multiple artefacts off a shared source-of-truth analysis (rather than independent prompts) embodies the "context as expertise" principle in a multi-output workflow