WHAT’S SHIPPED.

A stale second status badge on the job page could disagree with the live one beside it. There is one badge now and it is the live one.

An Auto job now goes from your brief through script, footage, voice and render without stopping for a Render click. When it is done, the editor is still there if you want to change something. Guided mode is retired. It was a second way of doing the same thing and it was splitting our attention. Existing Guided videos still open; you just cannot start a new one.

The script writer used to build a factual outline first and then write to it. That outline could seed a claim the sources did not support. It is gone: the writer now works from the source material directly, and the review pass sees the complete narration rather than a summary of it. If you supply your own script, it is now kept whole. It is no longer rewritten against the brief, the scene boundaries follow your paragraphs, and the video runs as long as your words take to say. The old word and duration limits on a supplied script have been deleted.

A production build had shipped with the development sign-in key baked in, so sign-in and sign-up were pointed at the wrong environment. That is fixed and the build now refuses to ship if it happens again. A subscriber asked for a copy of their data and the page sat on "Building your export" for hours. The export had failed underneath because the bundle was too large to pass through in one piece, and the failure never reached the page. Exports are now built in bounded pieces, a failure is shown as a failure, and a retry is safe.

A run of render failures from real jobs, each fixed at the cause. An image that a provider served as zero bytes took the whole final render down; it is now dropped and the next candidate takes the slot. One render chunk that never completed took the final down; it is now retried on a fresh connection and given a progress ceiling. A timeout drawing one look plate poisoned the rest of the render; it no longer does. And a step that could outrun its request and end the job with a 502 is now bounded. When a render does fail, you now get one clear failure, your credits back exactly once, and an email that says so. A retry can no longer overwrite a newer attempt. And a finished video lands under its own key, so a re-render never replaces the file you already downloaded.

The music bed now drops under the voice while someone is speaking and comes back up between lines, rather than sitting at one level throughout. We measured it on a real render rather than trusting the setting: the voice stays clear and the true peak stays under the ceiling.

An audit found that an annual subscription would have charged for a year but only granted the first month of credits, with nothing scheduled for the other eleven. Nobody was on an annual plan, so nobody was affected. Rather than leave that live, annual checkout is switched off and the annual option is gone from the pricing page. Monthly is unchanged. Annual comes back when the monthly credit schedule behind it exists.

The sample videos on the site have been re-rendered on the current pipeline. One that did not come back clean was pulled rather than left up.

For a period the look you picked was not reaching the rendered video. The step that draws the look plates had stopped working after a change to the render container, and nothing flagged it. It is fixed and every render since carries its look again. Sorry about that one. Every chunk of a long render was fading up from black for its first few frames, which showed as a flash at the seam. Fixed. Archival clips that open on a black leader now skip it. Smaller ones: film grain no longer washes the colour out with it, the colour description used for grading is one description end to end, a portrait shot is no longer cropped so the head is cut off, and a document on screen gets a proper zoom and highlight treatment.

When the narration names a year, that year now decides the era of the footage for that beat, even where the topic as a whole spans decades. Present-day briefs now ask for present-day footage instead of defaulting to archive. The referee that judges whether a candidate depicts the subject has been re-measured on a labelled set and moved to the model that scored best on it. The atmospheric fill used when nothing specific exists is now vetted the same way, and the variety candidates that stop a scene repeating are screened before they are offered. Each scene now carries a written visual contract, what the shot must show, and a candidate is judged against that rather than against the subject in general. A stock clip the referee rejected is no longer quietly rescued into the timeline further down.

Swapping a clip or regenerating a scene used to go straight out to the providers. It now reads the footage already found for your job first, and only asks the providers for the shortfall. The Suggested tab shows an "In this job" chip on those, and clips that were passed over only for file type or resolution are back in the list when you are the one choosing. Regenerating footage for a scene with no usable description used to spin forever. It now stops and tells you why, and it does so before any credits are charged. A matching fix underneath: a swap that could have matched two different candidates now refuses rather than guessing.

A long video with many scenes could run out of memory in the footage search step and fail before a single clip was chosen. The search now trims what it keeps in flight to a budget, scene by scene, so a long documentary gets through the same step a short one does. Alongside it, the archive lookups for a named person or place had no time limit and could hang a job indefinitely. Each now has one.

Several archives were feeding us less than they could. Openverse was being filtered by category, and that one filter discarded nine in ten of its results. The filter is gone. One stock source was throwing away anything it reported as below 1080p, a number we had never checked; that is gone too. And the size limit on archival posters was set to the wrong provider figure, so a slice of larger scans was being dropped before anyone looked at them. Medical and scientific history briefs can now reach the Wellcome Collection.

CueTheScene has been accepted onto the ElevenLabs Grants programme. Their logo now sits in the footer.

Share-alike licensed images are no longer used in comparison panels or evidence boards, because a share-alike licence would attach to your whole video. Everything in those slots now comes from sources whose licence you can publish under without conditions. The source labels shown in the credits now come from one list, so a provider is named the same way everywhere. And where a video includes an AI-generated intercut, that is now included in the disclosure text we prepare for YouTube.

The model that writes your script has been upgraded to a stronger one, and the pass that reviews the draft runs on a separate model so the writer is not marking its own work. The opening is now held to a hook inside the first 30 seconds. A comparison that the script cannot support is refused rather than kept, and a set of stale checks that could fail a good script for no reason is gone. One later fix: a script that opened with a year, "1993. A microbiologist...", was being rejected as a list and could fail the whole job. It no longer does.

Two things a history video must never do, both now blocked. Where we could not find a real photograph of a named person, the video could generate one. It looked like an archive still and it was not. That path is closed: a real person is shown from real sources or not shown at all. In the same pass, imagery of atrocities is no longer used as decoration in collages or grids, and the archive furniture that scanned collections carry, the colour bars, the catalogue cards, the mount, is no longer mistaken for the photograph. Side-by-side comparison cards could put the later moment on the left and the earlier on the right. Dates on those cards are now checked and the order is enforced, so a then-and-now reads then, then now. Diagrams in those slots keep their place instead of being sorted away.

Deleting your account already needed you to type a confirmation. It now also asks you to re-verify with your sign-in method first, so a left-open session cannot be used to remove an account and everything in it.

When a cancelled subscription reached the end of its last period, the final settlement from our billing provider was being treated as a renewal and a full month of credits was added to the account. Those credits were not paid for and would have vanished with the account. Fixed the same day it was found. The end of a cancelled subscription now grants nothing, and only a real renewal adds credits.

If you retried a video, the page kept showing the failed state until you reloaded it. Worse, a page you had left open never updated again at all, so a thumbnail that finished after the render never appeared and a video you published elsewhere never showed as published. Both fixed. A retried video now updates in place, and a page left open keeps receiving updates.

A ten minute video was coming out around 500MB, which is slow to download on a poor connection and slow to push to YouTube. We changed how the final file is encoded and the same video now comes out roughly half the size. We measured the quality either side rather than guessing: the difference is not visible at normal viewing, and the detailed, grainy, archival shots that are hardest to compress were the specific thing we checked.

The render step used to go out to a third party. It now runs on our own infrastructure, and that change is complete for every video. What you get from it: renders are quicker, the progress bar reflects real work rather than a spinner, and when something goes wrong in a render we can fix it the same day instead of filing a ticket with someone else. Two reliability fixes landed with it. Long videos were occasionally failing at the very last step, while the finished file was being assembled, and that is fixed. And if a render does fail and you retry it, it no longer starts again from nothing.

A group of editor changes. You can now favourite a voice so it floats to the top of the list, and jump straight to the videos you last used a voice on. You can block a footage source, a collection or a specific creator in settings, and we stop pulling from it. Useful if one archive keeps giving you material you do not want. And the editor stopped telling you things that were not true: re-voicing a scene now reports honestly when it is skipped or fails instead of sitting on a spinner, and a couple of actions that quietly did nothing now say so.

Picking a look, the grade and treatment applied over your footage, was changing what you saw in the editor but was never being applied to the rendered video. It had never worked on either render path. It works now. The look you choose is what comes out.

A continuation of the footage relevance work, from watching real videos back. One archive was serving oral history recordings and filmed lectures as if they were footage, so a documentary could end up showing a person sitting talking to camera in place of the thing being described. Those are now rejected. Archival clips that open on a black or blown-out frame, common on scanned film, now have that opening trimmed rather than showing it as the first thing on screen. Where an archive search came back thin, the video was falling back to holding one image for a very long time; the search now widens properly instead. And a caching fault that could re-add a picture a scene already had, so the same image appeared twice, is fixed.

Three things about the audio mix. Where there is a gap between narration, the music now lifts into it and settles back when the voice returns, instead of staying at one level through the whole video. Where a music track has to loop to cover a long video, the loop now crossfades instead of restarting, so there is no audible jump at the join. And sound effects are now set relative to the measured narration level rather than a fixed number, so they sit under the voice properly however loud your voiceover came out.

One speaking-rate number was being used for all ten voices, so the estimated length of a video drifted from the real one depending which voice you picked, and pacing was slightly off for most of them. Each voice now uses its own measured rate. Separately, the pass that reviews and improves a script was rejecting every single one of its own rewrites, so it was doing nothing at all. It works now.

You can now see exactly where your credits went: every charge and every refund, itemised, rather than a single balance that moves without explanation. Alongside it, the pay-as-you-go spending cap now resets with your billing cycle instead of drifting, and when you do hit it that is now visible rather than something you infer from things stopping.

The step that chooses between generated thumbnails was falling back to picking the first one whenever there were fewer than four to choose from, rather than actually comparing them. It now judges them properly however many there are.

Videos made before the editor rebuild could not be opened in it, because they were stored in a shape the new editor did not understand. Those are now converted on the way in, so an older video opens and edits like a new one.

Every video now carries a record of how it was put together: where each piece of footage came from, who made it, and the licence it was used under. It was all being stored already and never shown. This also fixed a real bug underneath it. Licence information was being discarded when footage was saved, so some attribution was being lost before it ever reached the finished video.

When a video puts a document, a letter or a newspaper on screen, it now understands what it is looking at rather than treating every picture the same. That means a book jacket is recognised as a book jacket instead of being placed as a document, pages with printed text baked into them stop being used where clean ones exist, and the pinned-evidence boards that true crime and investigation videos use can now pin an actual document rather than only faces.

Two visual faults at scene changes. A still image was ending a fraction of a frame before the next shot began, so some cuts flashed hard black for a single frame. The gap is gone. And where a clip did not fill the frame, the letterboxing around it was painting solid instead of letting the background through, so a shot that should have sat on your chosen backdrop sat in grey bars instead.

Side-by-side comparisons were choosing images badly. The picture for a person, a place and an object all went through the same generic search, so a comparison could end up showing something only loosely related to the thing being compared. Each now routes differently depending on what is being shown, and the candidate is checked against the subject before it is used. Where nothing suitable exists we generate an image rather than settling for a bad match. One small fix went with it: a date callout would not accept a year before 1500, so medieval and ancient subjects were silently losing their dates.

The music bed was mixed to a fixed loudness while the narration was not levelled at all, so on a quiet voiceover the music sat about a decibel under the voice and buried it. On some videos you could barely hear the narration under the bed. The voiceover is now levelled at source and the bed is set relative to the measured voice, so the music sits properly underneath it on every video regardless of how loud the voice came out.

Every scene change was holding a blank frame for a fraction of a second before the next shot appeared, because the transition faded out to background and back rather than overlapping the two shots. On a long video that is a visible flicker at every join. Fixed on both render paths.

The biggest step yet on footage relevance, from watching real videos back. When you name a person, we now check that the candidate footage genuinely depicts them and reject the ones that do not, instead of trusting a keyword match. That was the failure where a documentary about one scientist showed a photograph of a different, more famous one. We also stopped archive furniture being used as content: microfilm index cards, film resolution test charts and reel slates were being placed as if they were footage. Those are now recognised and rejected. This is a step, not a finish. Off-topic and wrong-period footage is still an open item on the roadmap and we are still working through it.

Captions, chapter cards, lower-thirds and callouts were rendering translucent straight over moving footage, so light or busy shots swallowed them. Every text family now sits on its own surface, sized to the text, so it reads over a pale document scan or a bright sky the same as over anything else. The same pass fixed the opposite failure: a handful of full-frame graphics were drawing their backing panel with nothing behind it, which showed up as several seconds of near-black carrying only text. Those now let the footage through.

Two things were making image-heavy videos, history especially, feel static. Detailed archival stills were being zoomed so far in that engravings and manuscript pages became unreadable, and single images were being held for 40 to 50 seconds with barely any movement. Detailed stills now get a gentle, restrained move that keeps them legible, while ordinary stock keeps its personality. And where the same image covers several beats it now reframes across them, opening wide, pushing into the detail worth seeing, then pulling back, the way an archive documentary actually handles a photograph.

Every video used to open on a static title panel showing a title the viewer had already read in the thumbnail, which spent the three most retention-critical seconds a YouTube video has on nothing. That is gone, along with the typographic hero opening that replaced it on some jobs. Videos now open on the first shot with narration from the first second.

A lot of good Wikimedia and Commons footage was being found and then quietly dropped, because the file could not be reliably fetched at render time. Those clips are now pulled in and converted up front, so archival video that used to fall out of the edit stays in it.

Two changes. Your thumbnail is now generated when the video parks in the editor rather than after the render, so it is on the dashboard, in your library and in the editor before you have spent anything on rendering. And they look different from each other now. Everything was coming out of four fixed styles on niche-locked palettes, which is why a history video and a tech video could look like the same thumbnail twice. The styles are wider, the per-video colour actually varies, and we moved to a stronger image model.

The progress page was doing too much. It has been rebuilt around one question: what stage is my video at, and how long left. One card, the stage you are actually on, and a time estimate measured from real jobs rather than guessed. The render leg shows real progress as each chunk lands instead of sitting on a spinner. When it finishes you get a single card with the video, the thumbnail and the way onward to publishing.

Videos no longer render themselves the moment the pipeline finishes. Every job now parks in the editor with everything built, so you can watch it back, swap footage, fix a caption or change a look, and then press Render when you are happy. Nothing is spent on the render until you say so. If you walk away mid-build we email you when it is parked and waiting.

You can now blocklist stock footage providers or specific collections on a per-channel basis. Any source you block will be skipped entirely when we pull footage, so you're not manually overriding the same picks every time.

You can now save your channel's niche, voice, look, and music mood as a brand profile. When you start a new video, load the profile and your preferences are already set. No more filling in the same options every time.

You can now shift-click to select a range of clips and move or delete them all at once. If you're reorganising a long timeline, this saves a lot of repetitive clicking.

You no longer need to pick a format or set a length before generating. Paste in a description of your video and we'll read it to decide whether a documentary, listicle, or another structure fits best, along with a sensible runtime. Less form-filling, faster first draft.

The projects list has been replaced with a card-based library. Each project shows a thumbnail, a status chip, and you can filter down to find an old video without scrolling through everything.

You can now apply colour grades directly in the editor, with options including black and white, vintage, and desaturated. Grades can be applied to a single clip or across the whole video in one go.

You can now browse a library of sound effects, whooshes, risers, and ambience, and drop them wherever you want on your timeline. No more hunting for audio files outside the tool.

You can now edit your voiceover, footage, music, and captions as separate lanes in a proper timeline, rather than working through a flat scene list. Each track is independently adjustable, so you can nudge timing on one element without touching the others.

Post requests and bugs, vote on what matters, watch things move across the roadmap, and see what shipped here. When we ship something you voted for, you get an email.

Scripts now open on a hook, carry retention beats through the body and end with a proper outro and subscribe prompt, tuned per format instead of stopping dead.

Footage now carries a niche-matched film-stock treatment: grain, grade, and a vignette on hero beats, instead of a plain pan. History looks period-authentic, science looks clean.

Hard cuts between shots instead of a soft dissolve on every join, and a cap so the same clip cannot be reused over and over across a video.

Chapter cards, lower-thirds, stat callouts, animated maps and charts now render at a proper density, a few every minute, opening on the hook. This was the single biggest thing making our output look flatter than it should.

The publish screen now walks the same steps Studio expects: details, elements, checks, visibility, with copy and download buttons inline at each step.

The landing page, the create screen and the in-app estimate now all say the same thing, measured from real jobs: about 20 minutes, 15 to 30 as the range.

One rendering path was occasionally leaving the thumbnail blank. Fixed, plus a nightly backstop that fills any that slip through.

You now get an email when a render is parked waiting for your review, and failure emails that say which stage broke, with the refund amount shown when one applies.

Simplified the states you see while a video builds, removed the confusing "your turn" messages on the automatic path, and put a retry button on every failure. If something stalls you can cancel and start again.

Some downloaded videos needed to buffer the whole file before they would scrub. Every render is now remuxed so it starts instantly and seeks cleanly, on both render paths.

We replaced the in-app YouTube upload with a Publish Kit: download the video, thumbnail and captions, copy a ready-made title, description and tags with the AI disclosure baked in, then open YouTube Studio. Fewer permissions to hand over, same end result.

We are open. Give us a topic and a format and you get back a finished video: script, voiceover, footage, captions and thumbnail, in about 20 minutes. Founding-member pricing is on while the early slots last.