# How to turn CoreSpeed into a content factory (/blog/how-to-turn-corespeed-into-a-content-factory)

We made PRODUCTION INCIDENT by taking a five-minute screenplay through Blender rehearsal, generated performances, recorded interfaces, and a finished 4K edit. CoreSpeed helped the agent retain production decisions and run media jobs. The work itself followed six stages, with an explicit review before each handoff.

[PRODUCTION INCIDENT | AI Short Film | 4K](https://www.youtube-nocookie.com/embed/rJ3-nvZu2M0)

## 1. Write actions before choosing shots. \[#1-write-actions-before-choosing-shots]

The film follows Maya and Alex through a remote call: a design preview fails, they investigate, restore missing files, and verify the result. The characters and incident are fictional. For every scene, we specified the immediate task, physical action, and resulting state.

“Investigate the failure” is difficult to perform. “Read the failed request, compare it with the release, then pause the rollout” gives the actor something concrete to do. Typing starts and stops for a reason; the next reaction follows a result.

We recorded the dialogue before locking the edit. Cut around completed actions, answers, and reactions rather than a fixed rhythm. Alex's admission stayed on his face through the question and reply. CoreSpeed memory carried approved constraints between sessions; the current script and shot plan remained the production reference.

## 2. Rehearse the complete sequence in Blender. \[#2-rehearse-the-complete-sequence-in-blender]

We built simple desks, computers, characters, and cameras, then cut the graybox against the actual dialogue. This exposed physical mistakes before final generation: a computer facing the wrong way, screen-poking gestures, and aimless keyboard activity.

The repair was specific blocking. Select a message, type a response, press Return once, stop, and read. Place the equipment so those actions are possible, then choose the camera. Check eyelines and hand positions when changing angles.

![A Blender graybox frame showing Maya at her workstation.](/blog/how-to-turn-corespeed-into-a-content-factory/blender-graybox.jpg)

Blender rehearsal: check the workstation, action, framing, and timing before generating final performances.

The landscape pass added profile, close-up, shoulder, hand, and reaction coverage, with selected slow pushes and short arcs. We reviewed whole sequences, because an attractive frame says little about whether the conversation feels rushed. The rehearsal established intent; generated footage still needed its own inspection.

## 3. Approve the cast and rooms from several angles. \[#3-approve-the-cast-and-rooms-from-several-angles]

We prepared environment, close-up, profile, and shoulder references for each character. One portrait could not establish how the face, headset, desk, and background should survive a change of angle.

Maya's warm design studio and Alex's cool home office made the remote locations distinct. Across each reference set, we checked clothing, hair, headset, microphone, device count, lighting direction, and room layout. After approving the night direction, earlier daylight footage stayed outside the final assembly.

![Maya's approved warm studio reference.](/blog/how-to-turn-corespeed-into-a-content-factory/maya-reference-preview.webp)

Maya: warm studio, green short sleeves, established headset and workstation.

![Alex's approved cool office reference.](/blog/how-to-turn-corespeed-into-a-content-factory/alex-reference-preview.webp)

Alex: cool office, navy overshirt, and a separate room.

## 4. Generate one directed performance at a time. \[#4-generate-one-directed-performance-at-a-time]

Each shot record linked a reference angle, action, camera instruction, source duration, and selected edit range. We tested difficult pilots before expanding the batch: a sustained close-up, a hand action, and a moving shot with a monitor.

The production used Gemini Omni for actor footage, Eleven v3 for dialogue, Sync v2 Pro for lip sync, and Stable Audio 3 Small Music for the score, through CoreSpeed's media tools. Saved job IDs let another session collect unfinished work without repeating the request.

A shortened close-up prompt illustrates the level of direction:

```text
Match Alex's approved close-up and cool night office. Preserve his
face, navy overshirt, headset, and desk layout. One continuous
eight-second take: listen, answer quietly, then listen again before
admitting the missed check. Keep his mouth still during the remote
colleague's lines. Make one slow camera push; no internal cut or
framing reset. No extra person, keyboard, captions, or invented
interface text. Exact voice timing and readable screens follow in post.
```

Review the returned performance at normal speed. Check the action, listening, eyeline, devices, and camera continuity. Inspect doubtful details at full resolution; a dark contact-sheet thumbnail once made an intact headset look missing.

![Contact sheet of Alex's generated performance at successive moments.](/blog/how-to-turn-corespeed-into-a-content-factory/generated-performance-contact-preview.webp)

Frames from Alex's generated eight-second pilot, reviewed before lip sync.

## 5. Finish speech and screens on the accepted take. \[#5-finish-speech-and-screens-on-the-accepted-take]

For lip sync, we supplied only the visible actor's dialogue, preserving silence during the other person's lines. The complete conversation could make a listener mouth the wrong words. Returned clips needed checks for truncated endings, timing, and visible listening. Generated clip audio was muted in the edit; the locked soundtrack remained authoritative.

Readable interfaces came from working browser recordings. The local demo actually saved a revision, failed a module import, restored files, and checked recovery. Product screens used actual components with fictional fixtures. Full-screen inserts were reserved for evidence the viewer needed to read.

![Two monitor inserts in an actual final film frame.](/blog/how-to-turn-corespeed-into-a-content-factory/screen-composite-preview.webp)

Real interface recordings composited into the accepted actor footage.

Track the display on the synchronized source: lip sync can change its geometry. Warp the recording into the monitor's four corners and preserve foreground objects. Our blue-lit headset earcup needed a corrected mask to stop interface content crossing it. Check those edges in motion and at native resolution.

![Coordinate grid over a headset and microphone edge for compositing inspection.](/blog/how-to-turn-corespeed-into-a-content-factory/screen-mask-review.jpg)

A close-up used to check the headset and microphone edges during screen compositing.

## 6. Assemble, mix, and inspect the encoded film. \[#6-assemble-mix-and-inspect-the-encoded-film]

HyperFrames kept footage, screens, captions, branding, and audio separately editable. Every shot retained its selected source interval. We removed distracting flashes between screen inserts and checked that actions and final syllables survived the cuts.

Sound began with keyboard taps and the Slack call alert. Three original music cues joined with four-second crossfades, without loops; music dropped beneath speech and thinned around Alex's admission. We kept purposeful keyboard and mouse effects, omitted environmental noise, and preserved the original speech speed.

The final master was 3840 × 2160 at 24 fps. A render problem taught us to place picture boundaries on integer frames while keeping audio at sample precision. We checked the complete encode technically, then inspected cuts, screen changes, captions, and difficult masks in the actual output. The handoff retained the script, references, accepted sources, soundtrack, screen recordings, edit project, and review evidence needed to make the next revision.

![Consecutive frames from the final encoded film used to review dialogue and captions.](/blog/how-to-turn-corespeed-into-a-content-factory/encoded-film-review-preview.webp)

Frames from the encoded film, checked for speech endings, captions, and continuity.