Showspring

Brainstorm → script → voice readout → location mapping → scene gen → video gen → trim → audio mix → render → publish. End-to-end on the VPS, no editor in the loop for the happy path.

Frame-accurate timeline accumulation across 30+ clips per episode. ffmpeg drifts when you concatenate floats, and on an 8-minute episode the audio walks off the picture by the back half. The pipeline now walks the timeline as integer frames and every consumer reads from the same canonical manifest.

YouTube publish had no idempotency. A retry after a network blip would post the same episode twice. Caught when an early episode hit the public feed in duplicate within 90 seconds; the publish step now keys off a content hash before posting.

The pipeline runs daily. Adding a new step inside this surface is a one-day lift. Adding a NEW production mode (Shorts, Podcast, Guest preview) costs multiple weeks — the cross-step contracts are tighter than they look. That tradeoff became visible in v1.1.

Per-step model picker on every creative surface. Local pool (Qwen3:32B / Qwen3:8B / Gemma3:27B / Gemma3:12B) plus cloud (Gemini, Claude). v2.1 extended this with a routing bridge that runs admin Gemini and Claude calls through local tooling when the GPU is online.

Mid-episode model swaps. The same project might brainstorm on cloud and script on local; the show-bible context has to load identically into both. The bridge layer normalizes input and output across providers so calling code never branches on model identity.

Quality dropped briefly when Idea Lab defaulted to Qwen 8B and the show-bible context spilled out of the window. Output went generic. The llama-swap router now picks the largest model that fits the prompt + still fits in VRAM with whatever else is warm.

Cost per episode dropped substantially once brainstorming defaulted to local. Quality stayed where it was. The "live routing indicator" in the topnav was added later in v2.1 because it was no longer obvious from the bill which engine had run which step.

Batch creation, per-batch scheduling, multi-platform publish (YouTube Shorts, TikTok, Reels, Facebook, X). Reused the v1.0 publish surface so the calendar and analytics work didn’t need to be re-implemented.

Per-clip image-to-video timing alignment. Each clip is independently rendered then concatenated, and clips with slightly off frame counts caused cumulative drift across the batch. Same fix as v1.0: integer-frame accumulation everywhere.

TikTok upload failed silently for two days because the token-rotation cron didn’t fire on weekends. Errors landed in a swallowed try/catch. Added a watchdog that pings each platform’s upload endpoint hourly and emails on failure.

Shorts now ship more reliably than episodes (smaller surface area, fewer moving parts). Sharing the publish/calendar surface across episode + shorts saved roughly two weeks of duplication that v1.0 had warned about.

7-step audio pipeline. Same script writer as video, different render path: per-character ElevenLabs voice synthesis, mix, podcast cover art generation, RSS publish.

Per-speaker volume balancing. Each character voice has different inherent loudness from ElevenLabs; concatenating without normalizing produced “listen on headphones” hot spots. ffmpeg loudnorm now runs per-clip before mix, with a final pass after.

RSS feed had per-episode artwork URLs that were absolute to the dev domain. First publish to Apple Podcasts threw 404s on every cover. Now resolved through a CDN-aliased path that’s identical across environments.

Quickly became the format for material that doesn’t need visuals (interviews, monologues, behind-the-scenes). The surgical isolation of “video paths” vs “audio paths” was easier than expected because v1.0 had already separated them at the manifest level.

Three-step Flow surface: server bundles start frame + prompt to Drive, operator generates manually, uploads MP4 back. Plus a dual-source image watcher (Drive polling + browser File System Access API) so iPad drawings and local PC folders feed the same approval queue.

Round-tripping arbitrary frame counts. Flow output isn’t always exactly 8.000s; the trim editor has to clip to integer-frame boundaries without re-encoding. -c:v copy with a -t flag computed from the manifest’s frame count works.

Drive polling stopped seeing new files because the OAuth token silently expired and the refresh path swallowed the failure. Two days of operator confusion before a watchdog caught it. Refresh path now logs and the topnav shows a Drive-disconnected indicator.

Became the v2.1 backbone for Guest Submissions and Shorts-from-Podcast: any time the right move is “a human generates this externally,” the Flow surface absorbs that without breaking the rest of the pipeline.

Three-persona debate (skeptic, optimist, story-finder) in parallel, judge model with live web search picks the winning angle. Selectable per episode.

Latency. Grok’s deep-research mode runs 60 seconds per call; three-way debate plus a research-grounded judge can take 6 minutes on news-heavy material. The UI streams each model’s progress so the operator can feel forward motion.

Grok kept hitting a 60s timeout in the bridge layer. Episodes silently fell back to a non-research path and the operator only noticed because output went generic. Bumped to 240s and added a UI banner when fallback fires.

Used selectively rather than by default. Most evergreen episodes don’t need it; news-cycle Gazette episodes lean on it heavily. The cost-quality tradeoff is per-episode.

Segment template system, character-positioned OTS graphics with live-preview sliders, intro videos per segment type, Discord auto-announce on episode publish.

4:3 OTS graphic composition over a 16:9 video frame. Anchor points had to stay tied to the character’s screen position even when the camera moved. Solved with a per-segment fixed-anchor map keyed to the character’s on-screen position rather than the absolute frame.

Cache-hash invalidation didn’t include the intro-video flag, so toggling it off didn’t re-render. The previous frame composition kept playing. Cache key now includes intro presence.

The Gazette format alone justified the v1.3 effort. Discord auto-announce was 2-day work and saved minutes per publish. The template system later grew to host more segment types without further plumbing.

Script-writer gate that decides when reactions are even appropriate. Multi-model planner that picks reaction TYPE (laugh, oh, hush, applause). Stem library with multiple takes per type. 9th “signoff” reaction that bleeds into the outro. Audio editor with lane-packed timelines so the operator can manually retime when needed.

Knowing when to react. The script doesn’t say “audience laughs here.” A small classifier reads each clip’s dialogue + tone and infers reaction type with timestamp. The non-trivial decision is silence: knowing when NOT to react is harder than picking which reaction.

Stem library kept hitting the same take in a row, which is the tell of a fake audience. Picker now tracks recently-used takes per type and avoids repeats within a window.

The audience layer makes a mediocre joke land. It also makes a flat one ROT, in slow motion. Operators learned to disable reactions on weak material rather than try to mask it with the audience.

Anti-slop hardening across every image-gen call. Burned-in TikTok-style large-bold captions. Top-level calendar with retention drill-down. Lock-for-editing on scheduled and published clips. Opt-in scheduler with confirm-on-publish.

The retention feedback loop. YouTube analytics aren’t real-time — there’s a 48-hour lag. The loop pulls retention curves into the calendar drill-down popover and uses them to flag scenes for re-cut. Required a separate analytics ingest plus popover positioning that survives screen edges.

A retention curve clipped through the popover boundary and bled onto the rest of the calendar. Cosmetic, but it kept slipping back across two refactors before the popover got a hard z-index seal and clip-path containment.

“Anti-slop” became a recurring lever. Every new feature now gets the question: does this make slop more or less likely? Sometimes the answer is to NOT ship it.

Five-level integration ladder (OTIO bundle → Lua importer → Python tray daemon → outbound WebSocket → Workflow Integration panel) on a single canonical JSON manifest. Frame-accurate timeline accumulation. ProRes 4444 stills. Markers carrying opaque Showspring identifiers for round-trip.

Volume duality. OTIO has no native audio-levels schema. Volumes ship on two channels: per-clip OTIO metadata embedded in the timeline, plus a volumes.json sidecar. Daemon applies via Resolve’s undocumented TimelineItem.SetProperty("Volume", dB) with the sidecar as recovery when SetProperty no-ops on certain Studio builds.

Windows Defender flagged the first daemon as a remote-control toolkit because PowerShell tray-icon code matched a heuristic. Rewrote in Python with pystray. Then Python’s stdlib HTTP server hung on Content-Length mismatches; switched to a leaner ASGI runner.

The OTIO contract is now load-bearing across the whole product. Round-trip via markers is infrastructure-ready — the next step is reading editor changes back into the database. Markers are stamped, Resolve’s GetMarkerByCustomData() works, the wire is up.

Two-pass LLM analysis (text first → characters and observations; photos second → personality summary with explicit additions-beyond-text delta + image descriptions + episode-idea seeds). ElevenLabs voice finder. Per-dog Google Flow video workflow. Speech-to-Speech voice replacement. Approvable public share link.

Voice-replace artifact compounding. Initial implementation read source audio from the most recent video, ran Speech-to-Speech, muxed the result back. Re-running compounded artifacts because each pass read its own previous output. Now reads from an immutable raw upload on every run.

The voice finder kept returning “no matches” on perfectly reasonable voice-direction prose because the LLM-extracted filters were over-restrictive. Added a progressive relaxation chain that successively widens until candidates appear; gender selector keeps the operator’s constraint hard.

Validates the v1.0 prediction: a NEW production mode is multiple weeks of work even when most components exist. Two-pass analysis schema, Drive folder bundling, ElevenLabs library cloning, Speech-to-Speech mux — each existed in fragments, none integrated.

Long-form podcast in, 9:16 vertical highlights out. Local ML pipeline does diarization, transcription, audio-visual active-speaker detection, and LLM highlight picking. Inline speaker assignment, configurable caption styling, lands in the standard publish pipeline.

Active-speaker detection that doesn’t lose the speaker on cuts. The first iteration used largest-face heuristic and dropped the speaker every time the camera switched. The audio-visual model now correlates lip movement with the audio track frame-by-frame; largest-face stays as fallback.

Render queue silently wedged on long jobs because a redispatch loop kept re-queuing the same job ID. Caught when an episode sat at 0% for two hours. Pipeline now keys idempotency on a content hash, not the request ID.

The Shorts publish pipeline absorbed the new clip type without re-implementing scheduling, calendar, or analytics. The v1.1 prediction (shared publish surface saves weeks) keeps paying out.

Three model legs returning calibrated 0–10 scores plus prose across five dimensions (growth, perception, traction, engagement, monetization). Radar chart for divergence, sparklines per dimension, per-dimension accordion grouping all three takes. Persisted as a historical trend.

Adaptive local model selection. The strategy call wants the largest reasoning model available, but the GPU is shared with active pipelines. Router checks which models are already resident in VRAM and picks the highest-priority warm model so the leg fires without a cold-load delay.

Reports landed inconsistent because two of the three legs would frequently time out on the first run after a cold start. Bumped Gemini timeout to 240s and added a streaming progress indicator so the operator can see the third leg arriving rather than assuming it failed.

Operator looks at the radar before greenlighting an episode batch — if engagement scores have been sliding three reports running, the next batch gets re-pitched rather than published. Replaces the static AI scoreboard the article used to host.

Three-tier bible: project show bible plus segment-specific bible plus a new segment_canon table with locked baseline rows (the Bonewagon, the parallel-tracks rule) and accumulated rows (eras / locations already visited). Dual-pitch flow extended to narrative segments — four destination + tonal-angle themes from Gemini Pro, eight pitches from Gemini and Claude in parallel, scored by Claude and Grok on five narrative dimensions. Operator picks a winning premise; a separate spine-build step generates the beat-by-beat content from that exact pitch.

Themes were prescriptive enough that two writers given the same theme converged on the same pitch — if the theme already named the genre, cast roles, and central event, there was nowhere to diverge. Rewrote the themes prompt so angle is tone/shape only, no cast roles or specific events, and gave each pitch writer an explicit “the other writer is drafting from the same theme — find an interpretation a careful reader of the theme would NOT predict” instruction. Operator now sees real choices instead of duplicates.

Pitch validators silently rejected valid Gemini Pro responses because the schema was strict on field names. Gemini sometimes returned summary or premise instead of concept under longer prompts. Cards stayed on “Writing…” while judges landed verdicts on a half-empty grid. Tolerant validator now accepts any concept-shaped field and normalizes; both attempts failing logs the actual JSON keys present so the next regression is one-pm2-grep away.

Operator-tagged characters are now hard cast across pitch and script — pitch validator runs a content-coverage pass (case-insensitive name search across concept, characterRoles, arc beats) and surfaces gaps; script writer post-pass enumerates clip character_ids and warns when a tagged character lands with zero speaking lines. Tagged cast no longer drops out between step 1 and the rendered episode.

Bridge concurrency upgrade: asyncio.Lock() → asyncio.Semaphore(4) for both Claude and Gemini CLI lanes. Four parallel CLI invocations per provider. Eleven previously cloud-routed call sites added to the bridge callsite map: narrative pitch / themes / verdicts, podcast cover description, music description, locations fallback, YouTube publish metadata, thumbnail prompts, approve-topics title, shorts-publish chunk fallback, and others. Public idea-share endpoints (anonymous reviewer sessions) intentionally stay cloud-routed.

A four-theme dual-pitch fanout sent four parallel Claude calls into a bridge that serialized them through a single global lock. Each call took ~100s on Claude CLI, so the operator waited ~8 minutes for the pitch grid to fill. Diagnostic was hard because the Gemini side returned in seconds — only Claude cards stayed on “Writing…” long enough to make the architecture limit visible.

Audit found 11 unregistered Claude/Gemini call sites burning Anthropic credits per operator action because their callsite strings weren’t in CALLSITE_MODEL_MAP. The bridge’s decide() returned { route: ‘cloud’ } for any unmapped callsite even when the bridge was healthy and the operator was admin-eligible. New callsites for narrative segments cloud-routed by default until they were registered — the calls then choked on a bridge-only model name (gemini-3-pro-preview) on the public API and returned null, which the strict validator silently rejected.

Dual-pitch wallclock dropped from ~9 minutes to ~3 minutes on a 4-theme fanout. Anthropic credit consumption on operator-session work dropped to near-zero outside of bridge-down periods — the cloud path is now strictly a fallback, not a silent default. Five apps share the bridge (Showspring, PodSplit, WillWin, Create Studio, Family Chat), so the parallelism upgrade and the routing audit both compound across the stack.