People decide, in one place
Every review action, whether a Discord button, a dashboard click or a Kitsu status, goes through one module and lands in Kitsu. There is no second database to fall out of sync.
A note on an item ("start two seconds later", "wrong face on the cover") becomes an automatic retake: Clipper re-cuts or redraws that item and sends it back for review.
Spend money only where it changes the answer
A cheap model reads everything and a strong model judges the shortlist. Paid stages refuse to rerun without --force. Cut keeps picks whose in and out points haven't changed, and covers are kept while their cut is unchanged.
Every paid stage takes --dry-run, which builds the prompts it would send and prints the estimated cost without calling anything. Every model call is logged with its token counts and price. clipper costs and the dashboard break spend down by stage, stream and provider.
Models and prompts are swappable and versioned
All model calls go through one layer, and the provider comes from the model string's prefix (Claude, OpenAI, Gemini, OpenRouter or local). Prompts are versioned files. A version that produced reviewed output is never edited; changes go into a new version.
clipper compare runs candidate models against a gold set of past reviewer decisions before a stage switches models.
One install, many shows
Config is layered: code defaults, a shared file, this machine's own file, then environment variables. Each show gets its own table with cast, style, budgets and layouts, and a single stream can override anything.
Adding a show is a config table and a Kitsu production, with no code changes. clipper doctor checks every dependency, key and service and prints the fix for each failure. clipper setup writes a new machine's config.
State that survives crashes
JSON state is written to a temp file, then swapped in. Updates take a lock, and readers retry when they catch a file mid-replace. Library code raises one error type with a one-line message, and only the CLI turns it into an exit.
Tests enforce the conventions: a test fails on a raw JSON read or a stray SystemExit, on a broken docs link, or on a CLI command missing from its group.
A schedule that stays current
Uploads fill fixed daily slots by type (Shorts, segments, compilations), alternate between shows, and stay under the channel's daily limit. Approved items older than a few days drop out, so the channel never falls weeks behind the shows.
The first version queued everything, and the newest stream ended up 15 days behind. Switching to "stay current" was a product decision, made from the data.
Measure before switching models
clipper compare runs a candidate model or prompt on streams that people have already reviewed, and counts how many of the approved moments it finds and how many windows it keeps. Running the same prompt twice first sets a noise floor, so a change is adopted only when it beats that.
Twelve models were tried through OpenRouter, with reasoning on and off. A newer restructured prompt found 8 of 18 approved moments against 6 for the previous one, but the gap is within the noise, so it stays a candidate. Every trial and its cost is written up on a wiki page.
It checks on itself
clipper checkup reads only the files the services write: stale heartbeats, a stage stuck for hours, a recording nobody picked up, an upload late for its slot, a title over YouTube's limits. clipper status opens with what needs a person, then puts services, recent streams, review counts, the next releases and the month's spend on one screen.
Housekeeping runs on its own too. clipper streams prune clears the source video of streams that are finished, never anything still waiting, scheduled or pinned. clipper backup zips the parts made by hand (config, schedule, enrolled faces, voices, layouts, cast photos, review state) every week, without secrets.
A nightly report is built on top of it. When nothing is wrong it sends a one-line DM with no model call. When something is wrong, Claude Code runs headless with read-only tools and sends one DM with its diagnosis. If that fails, the raw findings go out instead. A weekly digest covers runs, uploads, spend and review decisions, and once a month a read-only Claude pass reads those weeks to find the problems that keep coming back.
A bug worth telling
The compilation that repeated itself
A reviewer noticed a speaker "repeating himself" two hours into a compilation. Lossless cuts have to start on a keyframe, so each segment file began up to five seconds before its real start. Joined back to back, every join replayed the tail of the previous part: 13 of 15 joins in that video.
The fix ends each part with a concat outpoint exactly where the next file begins. The cuts stay lossless and the joins are now frame-accurate. A regression test covers it.