What the proposal covers.
The final preview was compared with the supplied discussion screenshot. These notes map the ideas to a feasible direction and show what remains unproven. The browser does not run AI, record a microphone or modify Premiere.
| Discussion idea | Proposed approach | Status |
|---|---|---|
| Conversation, hooks and brainstorming | Local model + selected transcript + FTF guidance. Scripted in this preview. | Feasible new work |
| Voice input | Push-to-talk → existing local speech engine → editable text → reviewed request. Mic stays off in this demo. | Feasible new work |
| Knowledge of FTF lessons | Retrieve approved lesson text, recipes and examples. Retrieval is not proof of expert-level training. | Feasible new work |
| Three ten-second preview choices | Compare a short excerpt before a larger edit. Current visual treatments use the same footage; distinct pacing, music and native graphics need rendering work. | Needs proof |
| Human check-ins and customization | Refine a selection, keep cuts locked, review exact changes, approve and undo. The Premiere connection must enforce these limits. | Needs proof |
| Talking-head rough cut | Reuse cached transcript and existing Smart Cut engine through a tested Premiere connection. | Needs native proof |
| B-roll guidance and own-footage scope | Suggest shot briefs and missing-footage markers. Search only selected owned media when requested; a marker is not inserted footage. | Feasible new work |
| Music, overlays, effects and LUTs | Browse and audition assets, then use explicitly supported product actions. Download/import is different from applying or replacing effects. | Mixed: source exists; adapters need proof |
| Dynamic music/asset replacement | Target only the requested component and preserve other tracks. Real-time speed is unmeasured. | Needs native proof |
| Cinematic Titles | 13 native Premiere title transitions. The demo changes an illustrative existing instance’s accent, not a native rendered title. | Product exists; connection needs proof |
| Frame Guide | Optional composition overlays; not required for the first Smart Cut proof. No automatic crop or tracking claim. | Product exists; connection needs proof |
| FTF asset advantage | Reuse a shared catalog, separate product/pack rights and a versioned local cache. | Existing source + unfinished access link |
| Popular assets and new-content notices | Opt-in aggregate usage and versioned catalog releases. No fabricated popularity or guaranteed weekly schedule. | Feasible new work |
| Quality over quantity | Improve recommendations while retaining asset versions referenced by projects. | Feasible new work |
| Learn from prompts and choices | Separate local preferences, analytics consent and explicitly contributed training examples. No automatic upload or continuous retraining. | Feasible new work |
| Training examples and feedback | Use owned examples, explanations and rejected alternatives; evaluate before fine-tuning. No fixed data ratio or training timeline is established. | Requires an approved dataset |
| Free local, paid API or owned server | Local is the default proposal. API/server options are separate, optional and metered; their displayed rates are example inputs. | Architecture options |
| Pricing, usage and margins | Calculators estimate selected costs. Subscription prices, included usage, profit margins and build credits need actual measurements and business decisions. | Not approved pricing |
| 90% finished / four prompts / launch and beta targets | Product ambitions, not demonstrated performance or launch commitments. | Unvalidated goals |
| Changing songs and copyright | Use owned/licensed music. A fixed amount of change does not automatically establish permission or new ownership. | Rights decision |
| Other tools and launch examples | Inspiration, not a required paid API or proof of our implementation. The referenced numbered points 14/16 were not supplied. | Context / unresolved reference |
Sample footage
The ten-second preview comes from the FTF webinar file supplied for this preview, from 00:05 to 00:15. Existing video titles and subtitles are part of that source. The example cut words are separate sample text, not a transcript of this excerpt.
Product access and delivery
Creator Hub handles identity and independent product rights. Existing Hub, Frame Guide and Titles releases are separate from the isolated new asset-service candidate. The new session/access bridge and private asset delivery remain unfinished.
Current Dropbox shared URLs remain public and redistributable. A proposed private route publishes reviewed asset versions into protected storage, checks product/pack access and downloads only selected files. Cached files are not proof of ownership. Preserve project-referenced media when accounts change.
The latest recorded CineStems delivery has 14 full mixes, 60 stems and 3 PDFs: 4,291,918,611 uncompressed bytes. The dynamic compressed Dropbox ZIP size is unknown. One selected full mix is 52.95 MB; its mix plus five stems is 317.72 MB. This review uses recorded delivery evidence, not a new live Dropbox test.
Native products stay independent
Cinematic Titles contains 13 native title transitions, not a MOGRT collection. Its visual demonstration models one accent color/amount change to an existing example. Native parameter access, faithful rendering and undo/save/reopen need their own proof. Manual Effect Controls remain available.
Frame Guide has existing PNG-overlay placement and opacity behavior. It is optional for the assistant and the first editing proof. The browser guide drawing is not a native integration test.
Open-source choices, size and costs
Qwen3-4B Q4_K_M through llama.cpp is a benchmark candidate: roughly 2.5 GB of optional weights, plus runtime and working memory. First reuse existing speech/transcript tools. A separate whisper.cpp base model is 142 MiB, or small is 466 MiB, only if justified. Local inference has no AI-provider usage charge; it still consumes customer hardware and electricity.
Existing source folders measured separately: asset candidate approximately 302 MB and Smart Cut runtime approximately 1.36 GB. These are not combined installer requirements. FTF course/index size is unmeasured. Three ten-second previews at 8 Mbps imply about 30 MB of media payload, excluding overhead and intermediates.
R2 storage examples use published Standard rates and an otherwise unused free allowance. The calculator excludes request overages, databases, other services, tax and development. API token rates and GPU-hour rates are editable examples, not provider quotes. Source pricing was checked September 29, 2026; verify before purchasing.
The next proof
After separate implementation approval: one transcript-selected removal on a duplicate Premiere sequence, reusing existing local tools. Validate exact frame mapping, protected tracks, rejecting an outdated review, A/V sync, undo, recovery and save/reopen. No new model, Frame Guide or backend is required for this first proof. Then prove one music-only operation, followed by local conversation/voice and richer previews.
More ways to use the plugin
Explore contains 48 task ideas across story, pacing, dialogue, captions, music, visuals, workflow and delivery. Each has a sample request and a connection path. Saving a request only changes this browser session; these are not 48 finished Premiere features.
| Task family | How we could build it | What needs checking |
|---|---|---|
| Story and editing plans | An optional local model reads the selected transcript and approved FTF guidance to draft openings, outlines, shot lists and notes. | Suggestion quality, hardware speed and human review. No new model is needed for the first cut test. |
| Pacing and transcript cuts | Reuse Smart Cut’s existing tools, then review the exact selection before changing a duplicate sequence. | Frame mapping, protected tracks, outdated reviews, sync, undo and reopening in Premiere. |
| Audio and render checks | Local FFmpeg filters can measure loudness, detect candidate quiet, black or frozen sections, and render separate comparison files. | Thresholds need tuning. Intentional silence and still frames are not errors. Separate renders do not prove native Premiere effects. |
| Captions | Use cached word timing for SRT drafts, line breaks, timing checks and reading-speed reports. A local model could draft translations for review. | Timing and language quality. Native styling needs its own connection. Adobe’s built-in caption translation uses online services. |
| Music, overlays and titles | Search owned catalog tags, audition files, and use only the actions each product supports. | Product rights and private delivery still need connecting. Stem suggestions do not guarantee every mix contains the requested instrument. |
| Shot markers and framing | Adobe documents marker actions. A shot brief can become a reviewed marker. A manual crop preview can show framing choices. | Support differs between host versions and bridges. No automatic subject tracking or generated missing footage is claimed. |
| Project organization and handoff | Read file metadata and create inventories, size estimates, version reports and checklists. Hash files only when duplicate checks need it. | Reports do not move or delete media. Renaming, moving or changing bins needs a separately supported operation. |
The first test stays small: one approved transcript cut on a duplicate sequence. Add task families only after their own results are checked. Voice cloning, automatic retraining and complete autonomous edits are outside this proposal.
Primary sources
- FFmpeg audio and video filters ↗
- Adobe marker actions ↗
- Adobe caption translation and online processing ↗
- Adobe native editing actions ↗
- Adobe transaction guidance ↗
- llama.cpp — MIT ↗
- Official Qwen3-4B files — Apache-2.0 ↗
- Existing faster-whisper engine ↗
- Alternative whisper.cpp sizes ↗
- SQLite local text retrieval ↗
- FFmpeg licensing ↗
- R2 storage prices ↗
- Workers prices ↗
- Dropbox shared-link limits ↗
- US Copyright Office guidance ↗
