Supplement: The Realtor Reel Pipeline, Twenty-Five AI Filmmaking Moves, and Five Editors Graded Out of Ten

Five steps, each with one tool. ChatGPT brainstorms the angle - five mistakes people make buying a home - narrows it and writes the script. Nano Banana Pro generates the presenter and the B-roll stills at 9:16 and 2K, three variations a go. Kling 3.0 turns each still into a five-second clip, and the prompt has to name the camera movement or the result drifts. Narration is recorded once, then split in CapCut into seven scene-length WAV clips using in and out points, and each clip goes to HeyGen to produce a lip-synced talking head at 9:16, 1080p. Everything assembles in free desktop CapCut at 24fps: B-roll layered over the talking head, mismatched footage resized, auto-captions styled with an outline in bold Montserrat, yellow, sitting under the chin. Most of the models are reached through PromptEdit.com, which is pay-as-you-go rather than another subscription.

The failure mode of AI-generated video is that the presenter changes face and wardrobe between shots. Nick's fix is mechanical: upload the reference images into Nano Banana Pro and refer to them by number inside the prompt - image one is the person, image two is the setting, and so on - so identity and clothing carry across every generated frame. It is the same discipline that makes the first-frame/last-frame technique work later: give the model a fixed anchor rather than a description and hope. Everything else in the generation settings is housekeeping - 9:16 for vertical, 2K, three variations so there is something to choose between.

Generated footage costs money per second, and the two habits Nick closes on are the difference between a sustainable workflow and a bill. First, keep every piece of B-roll you generate - within a niche the same shots come round again, and reuse costs nothing. Second, the expensive generation is the lip-synced talking head, so produce one only for the intro and the outro and cover everything in between with B-roll you already own. Two smaller savings sit alongside: turning off audio generation on video clips, and using PromptEdit's pay-as-you-go pricing instead of holding a subscription to every model.

The 'twenty-five ways' lesson is mostly one method wearing different hats. Generate or export a still for the first frame of the shot and a still for the last frame, hand both to Seedance 2.0 on PromptEdit, and the model interpolates the motion. That covers transitions between two real clips that no camera could physically make; camera moves and speed ramps that would need a gimbal; ageing and de-ageing, using a normal photo as the start and an AI-aged one as the end; an infographic that draws itself, using a blank version as the first frame and the finished version as the last; and a day-to-night time-lapse generated from a single photograph. The sibling technique is motion reference - give it a still plus a reference video whose movement it should copy, which is how a light stand becomes a lightsaber.

Beyond the interpolation trick, the lesson is a catalogue of small post jobs that used to need a specialist. Unwanted objects come out by drawing an arrow at them in the reference image and asking for removal - and it works on moving shots too. Products swap brand or colour; wardrobe changes colour; whole backgrounds become a waterfall or a cafe; weather is stacked as timeline layers; the presenter plays both halves of a two-person scene. Voice is its own group: cloning through PromptEdit, but dubbing into other languages has to go to ElevenLabs because PromptEdit does not do it. Music and sound effects are generated by writing the brief in ChatGPT and pasting it into PromptEdit's generative tab. Finally Topaz Video upscales what is too small to use - iPhone 6 footage and low-resolution generated drone shots both demonstrated.

Across five graded edits the thing that most reliably moved the score was B-roll: whether it was chosen to serve the sentence being spoken, and whether the editor went looking for more. The top scorer sourced extra matching footage from the channel's own back catalogue, unprompted, and timed music to punch on emphasised lines and drop away on breaths - behaviour the host called that of an 'ultra, ultra skilled editor'. The 7/8 editor had the best graphics and sound design of the group and still lost points: 'B roll is crucial to that... I don't think this person hit it.' The inverse holds too - a genuinely good editor 'usually doesn't even need to lean in on effects and transitions and big graphics' if the footage choices are right.

The host runs an informal internal filter on every visual choice, and states it as a single question: 'I always use the term, would Apple use this in their marketing?' Comic-book rotoscope effects, money graphics, ticking clocks, gold and confetti all fail it. His objection is not taste, it is trust - crypto-influencer visuals make him look untrustworthy on camera and pull against a brand built on being clean and credible. One otherwise capable edit was rated 'a 10 on the cheese factor' and heavily marked down for exactly this, and cheesiness cost points even where the graphics and animation craft was among the best in the group.

The most specific technical note in the whole critique is about easing. Zooms, transitions and graphic movements set to linear give a video an amateur feel that most viewers register without being able to name; eased motion - ease in, ease out - reads as 'much more smooth and natural'. The host calls the fix 'very small but easy' relative to how much it changes the perceived quality of a brand. The same attention-to-detail family covers parenting: in a later edit an app icon animated smoothly while its own text label moved and scaled differently, because the two were never parented together.

One applicant's edit shipped with the blue warp-stabiliser 'needs reanalysing' bar still burned into a shot, plus a small unrendered glitch, and an earlier one left a stray mask edge visible in a graphic. The host reads all three the same way: the editor never watched their own export before sending it. In a hiring context he treats that as character rather than craft - it is 'like misspelling something on your resume' - and he scored that edit a 5 where his co-reviewer, judging pure editing skill, gave it an 8. His own remedy is to step away and take 'a quick walk' before the final export, so the second pass is done with fresh eyes.

No applicant was given brand guidelines. The host did that deliberately, to see who would research the company and infer its look. One rebuilt from scratch a background graphic that matched one the host had made himself and had never shown anyone, and matched the fonts and colours besides: 'I do really appreciate how even the fonts, a lot of the colors, everything really matched our brand style. Which just shows to me that this person cares.' The inverse is marked down - editors who impose a 'cookie cutter' house style over a client's existing identity, and, in two separate edits, editors who skipped the company's own signature logo animation.

The admin lesson carries five transferable rules for any peer community. Search the archive before asking, because the question has usually been answered. Give as much as you take rather than only extracting. Ask for feedback strategically and reciprocally, instead of dropping a link and hoping. Read and follow the rules - low-effort 'check this out' posts get ignored, specific questions get answered. And attend the live calls, joining early, because that is where the unrecorded value is. The platform detail is incidental but explains the design: the community left Facebook groups because the algorithm buried posts and spam crept in, moving to a chronological feed with an event calendar and a leaderboard.