Producing the Menu-Makeover Screencast Episode
Adding one narrated build episode to Gantry's screencast series
Gantry keeps a series of narrated screencasts of its own runs — short videos where you watch it build something, with a voice-over that tells the story sprint by sprint. Each new run that earns an episode has to be produced through the same pipeline: register it in the series, write a narration script anchored to what actually happened, record the voice-over, and render the finished video with the narration synced to the replay.
This job produced one such episode, for an earlier run that had redesigned Gantry's terminal UI into a menu bar with a pager, clipboard support, and live statistics. That run had already happened; what did not yet exist was its place in the screencast series. The narration audio for it had been recorded by hand — a paid text-to-speech step done outside the automated build — so the delicate part was not inventing the episode but assembling it honestly around artifacts that already existed and could not be casually regenerated.
Build
The pipeline was cut into curate, script, narrate, and render, but the load sat on the two middle pieces that were framed as verification of already-recorded, audio-aligned artifacts rather than as fresh generation.
The episode work divides along the production stages: register the run in the series and refresh the derived inputs, settle the narration script, settle the narration audio, then compile the timing and render the video. The ordering is what lets it work — nothing downstream has a stable episode entry to build against until the run is curated in with real replay provenance, so that piece goes first and the render, which turns everything into the shippable video, goes last.
The unusual part of the cut is that the two middle pieces were written as verification, not authoring. The script and the narration pair were already committed, and the narration had been recorded against that exact script text through a paid voice provider. The briefs say so bluntly: do not rewrite the script, do not call any speech synthesizer, do not overwrite the recorded audio — regenerating either would desync them and throw away a paid, hand-crossed boundary. So the decomposition asked those sprints only to confirm the frozen artifacts still pass their validators.
That is where the run came under stress. The script sprint — the one meant to be a near no-op — went red at the gate twice, took a fix, and cycled through investigation before it landed green, while the curation, narration, and render sprints passed cleanly. The difficulty in this build was not composing narration; it was getting the surrounding checks, including the cargo build the gate runs, to agree that the untouched, alignment-fragile artifacts were valid. No boundary was redrawn — the pieces held their shape — but the gate located the hard part precisely where the plan had tried to make it safe.
Feature
Before this job the screencast series had no episode for the menu-bar run; it now carries a curated manifest entry that drives the full narrated pipeline end to end, episode-scoped so rebuilding it leaves the rest of the series alone.
The series is how someone watches Gantry build, and one finished run was missing from it entirely — no manifest entry, so no script, no render, no way for the pipeline to treat it as a real replay. This job supplies the whole chain for that one episode: an included entry in the append-only episode manifest pointing at the immutable source run with its real timing provenance, refreshed series digest and map inputs so the episode joins the normal flow, the anchored script, the recorded narration audio and its alignment data, a compiled sync profile, the narrated video, and the upload-metadata sidecar. Producing it also touched Gantry's engine — the manifest digest, the shape derivation, and the sync-profile compiler — so the code that assembles episodes could handle this one.
Standing in the tree at today's HEAD, the entry is still there: the manifest lists the menu-makeover episode as included, pointed at the same source run, and every file the job introduced is still present — the script, the narration alignment data, the sync profile, the rendered video, and the metadata sidecar all exist where the pipeline expects them. The engine changes remain in place, and the bulk of the lines the job added survive unchanged. The rerender path stayed episode-scoped as intended: this episode's artifacts sit on their own numbered track without having disturbed the other episodes around them.