Agent Mode
Agent Mode is an AI-powered editing assistant built into modelBridge. Instead of navigating menus, adjusting parameters, and manually checking your work, you describe what you want in plain language — and the agent handles it directly in Premiere Pro.
Every editor works differently. There’s no single right way to check a timeline, prep for delivery, or organize a project. Agent Mode adapts to how you work — your standards, your shortcuts, your way of catching problems. Shape it with custom instructions, and it becomes your assistant, not a generic one.
Data & privacy
Section titled “Data & privacy”Cleanup features — silence removal, cuts, ripple-delete, timeline scans — run on your machine. Claude only receives timestamps and clip names to reason about what to change. Your video file never leaves your computer for this work.
When you want Claude to actually look at your footage (visual scan), a handful of small sampled keyframes go to Anthropic — never the full clip. That feature is opt-in per action.
For a full breakdown of what runs local vs cloud, see Privacy Controls. For NDA workflows, see the NDA editing guide.
Real workflows
Section titled “Real workflows”Dailies prep
Section titled “Dailies prep”You come home after a shoot with 80 raw clips. Normally 45 minutes of sorting footage into folders, renaming clips, setting color labels and status tags — mechanical work that has to happen before the creative can start.
User: “Prep today’s footage”
Agent: Reads every clip, identifies cameras from file metadata, groups clips by source, flags short takes and generated files for review, and proposes a complete organization plan in one message. You confirm or correct in one reply. Five minutes. Nothing to do tomorrow morning.
The metadata travels with the footage. Your colorist in Resolve sees the same camera labels — you didn’t have to tell them anything.
Timeline scan before delivery
Section titled “Timeline scan before delivery”You’re done with a cut and delivering tomorrow. Are there offline clips? Frame rate mismatches? Gaps that will export as black? Normally you scrub through manually and hope you catch everything.
User: “Scan my timeline”
Agent: Runs a 21-point inspection covering technical compliance, media analysis, and editorial polish. Findings grouped by severity: critical issues first, then warnings, then observations. Every finding includes the clip name and exact timecode.
Silence removal
Section titled “Silence removal”You have a 20-minute interview with long pauses and filler words. Normally you listen through and cut by hand.
User: “Remove silence from the interview track”
Agent: Measures the actual noise floor of your recording, shows you exactly how many segments it found and how much time will be removed, and waits for your confirmation before cutting anything.
Rough cut from markers
Section titled “Rough cut from markers”You watch the rushes once, marking every take that catches your eye. Now you have 30 markers across an hour of footage and the slow job of slicing out the keepers and stacking them in order.
User: “Build a rough cut from my markers”
Agent: Splits the timeline at each marker boundary, keeps the marked sections, and removes the rest in one pass. The cut you described while watching is the cut you get back.
Color consistency across deliverables
Section titled “Color consistency across deliverables”You have a master cut and three social versions. You’re not sure the color grade matches across all of them.
User: “Do all my cutdowns match the master’s color grade?”
Agent: Scans every sequence and reports which clips have a different LUT or no LUT at all, with clip names and sequence locations.
Understand a model
Section titled “Understand a model”You see a model in Browse but don’t understand what the parameters do.
User: Clicks “Explain this model” on a model card.
Agent: Gives you a short, honest walkthrough of what the model is best for in a Premiere workflow and which parameters actually matter. Say yes and it installs the model directly.
Cover the missing angle
Section titled “Cover the missing angle”You’re cutting a two-person dialogue scene and you only have one camera. The cut works for a while — then you need to break up a long beat and realise there’s no reverse shot to cut to.
User: “Build me a reverse shot of the second speaker that matches this frame” — with a frame from the timeline attached.
Agent: Reads the reference frame, identifies what needs to stay locked (face, wardrobe, set, lighting), and writes a prompt structured for coverage — correct lens feel, eye-level framing, identity preservation, scene-matched lighting and grade. Asks if you want it bound to a fal.ai model in the Generate tab. Say yes — the prompt lands ready to run with the reference frame pre-attached, one click to generate.
What it can do
Section titled “What it can do”At a glance — every capability and what it saves you
Section titled “At a glance — every capability and what it saves you”Browse the full list of what the agent can do for editors. Group by purpose, ask for any of these in plain language.
Read your project
Section titled “Read your project”| What you can ask | What you get back |
|---|---|
| ”What’s in this project?” | One snapshot of every bin, every sequence, the media types, offline items, and active sequence — instead of clicking through bins and sequence tabs |
| ”What’s in the Footage bin?” | Detailed item-by-item drill-down: media paths, offline status, frame rate, in/out points, color labels |
| ”What clips are on the timeline?” | Every clip across every track with name, dimensions, source frame rate, position, and duration |
| ”What’s selected?” | Only the clips currently selected — lighter and faster than reading the whole timeline |
| ”What effects are on this clip?” | Full effect stack, Motion properties, opacity, blend mode, speed, reverse state, disabled flag |
| ”Where are the gaps on V1?” | Gap location, duration, and the names of neighboring clips |
| ”What does the audio track layout look like?” | Track count, clip count per track, clip names and positions |
| ”Where’s the playhead right now?” | Current playhead position in seconds and timecode |
| ”What markers are on this sequence?” | All markers with name, position, color, type, and comments |
| ”What’s the track state?” | Every video and audio track with name, mute state, lock state, clip count |
| ”Tell me everything about this clip” | Exhaustive dump: every effect with every property value, source media frame rate, alpha, proxy status, XMP metadata, clip markers, speed, enabled state — in one call |
| ”Tell me everything about this sequence” | Exhaustive sequence dump: settings, every track with lock/mute/target, every clip with positions and effects, every marker, every transition, work area, in/out |
| ”What properties does this effect have?” | All properties on one effect with current values — discover what’s available before setting |
| ”Read the metadata on this bin item” | Project metadata + XMP for one item |
| ”What workspaces do I have?” / “What effects are available?” | Discovery calls so the agent uses valid names before acting |
| ”What’s loaded in the Source Monitor?” | Current item, in/out points, media path |
Catch problems before delivery
Section titled “Catch problems before delivery”| What you can ask | What you get back |
|---|---|
| ”Scan my timeline” | Inspects the timeline across up to 21 dimensions. Technical checks are measured directly from the media and project state (FPS, resolution, audio coverage, codec/sample-rate/channel mixing, bitrate outliers, LUT consistency, proxies, gaps, flash frames, disabled clips). Editorial observations (long static shots, repeated beats, heavy pacing, abrupt endings, missing B-roll) are the agent’s read of the timeline structure — surfaced for your judgement, not a guaranteed detector. |
| ”What codecs are in this timeline?” | Real codec/bitrate/sample-rate readout from the media files themselves — data Premiere’s API doesn’t expose |
| ”Which clips don’t have proxies?” | Project-wide proxy audit prioritized by what’s on the active timeline |
| ”Are any files offline?” | Every offline item with its expected path, ready to be relinked |
| ”Where are the silent parts?” | Noise-floor-calibrated silence detection with red markers for review before any cut |
| ”What’s a normal noise threshold for this clip?” | The agent measures from a quiet point on your clip — your environment, not a guess |
| ”Find where the host is alone in this interview clip” | Frame-by-frame visual analysis returning the time-ranges that match. Flags any range it couldn’t analyze and offers to re-scan; never concludes on data it didn’t see. |
Edit on the timeline
Section titled “Edit on the timeline”| What you can ask | What you get back |
|---|---|
| ”Move this clip to 00:01:30” | Clip relocated and verified back |
| ”Trim the end of this clip by 2 seconds” | New out-point applied — source-based or timeline-based |
| ”Split at every 5 seconds” | Razor cuts placed across the whole track in one pass |
| ”Delete this clip” (or “and close the gap”) | Either lift (leaves gap) or ripple delete (closes gap) — your choice |
| ”Enable/disable this clip” | Clip toggled without removing |
| ”Scale this to fill the frame” | Motion → Scale set to fit-to-fill |
| ”Move this to the top right” | Motion → Position adjusted |
| ”Half-speed this clip, reverse it” | Speed change and reverse — separate operations or combined |
| ”Set opacity to 50%“ | Opacity property updated |
| ”Reinterpret this clip as 23.976” | Frame rate interpretation changed across all timeline instances of that source |
| ”Set the in/out points on the source monitor” | Source-monitor in/out positioned for insert/overwrite |
| ”Move the playhead to the next marker” | Playhead positioned at a specific time |
Precision trim moves (the senior-editor toolkit)
Section titled “Precision trim moves (the senior-editor toolkit)”| What you can ask | What you get back |
|---|---|
| ”Ripple delete this clip” | Clip removed AND every later clip slides left to close the gap |
| ”Roll the cut between these two clips by 1 second to the right” | The edit point moves — one clip extends, the other shortens, total sequence duration unchanged |
| ”Slide this clip half a second later” | Clip moves on the timeline; neighbors adjust to compensate — clip duration unchanged |
| ”Slip this clip back by a second” | Source in/out points shift; the clip stays in place but shows a different part of the take |
| ”Reverse this clip” | Playback direction flipped, independent of speed |
| ”Turn on frame blending for this slow-motion clip” | Frame blending toggled for smoother slow-mo |
| ”Detect scene cuts in this clip” | Premiere analyzes the footage and auto-adds cuts at detected scene boundaries — perfect for chopping raw long takes |
| ”Strip every effect off this clip” | All effects removed in one call instead of one at a time |
| ”Rename this clip to ‘Hero Shot‘“ | Timeline clip renamed (source bin item untouched) |
| “Move this clip to V3” | Clip jumps to a different track |
| ”Undo the last 3 things” | Multi-step undo from the chat — best-effort: scripted edits don’t always reverse cleanly (see What it cannot do) |
Effects, color, and transitions
Section titled “Effects, color, and transitions”| What you can ask | What you get back |
|---|---|
| ”Add Lumetri Color to this clip” | Effect applied by name |
| ”Add a cross dissolve between these two clips” | Transition placed at the edit point |
| ”Remove the Lumetri effect” | Specific effect cleared |
| ”Set exposure to -0.3 on this clip” | Lumetri parameter updated — exposure, contrast, highlights, shadows, whites, blacks, temperature, tint, saturation, vibrance, sharpening, all by name |
| ”What LUT is on the hero shot?” | Input LUT name and path — or confirmation that there isn’t one |
| ”Do all my cutdowns match the master’s color grade?” | Cross-sequence LUT comparison with named outliers |
| ”Copy the LUT from this clip to everything on V1” | Batch LUT application with per-clip success report |
| ”Copy the Lumetri grade from this clip to those four” | All non-intrinsic effects propagated from source to multiple targets |
| ”Both clips have Lumetri — sync the exposure and contrast” | Property values copied between two clips that share the same effect |
| ”Apply Gaussian Blur to every selected clip” | Batch effect application — selection / one track / all clips in sequence |
| ”Set blend mode to Multiply” | Any standard blend mode by name (Multiply, Screen, Overlay, …) |
| ”Set exposure to -0.3 (no animation)“ | Static value on any effect property — for one-off settings without keyframes |
| ”What effects are available in this version of Premiere?” | Effect catalog — the agent uses this to verify names before applying |
Animate over time (keyframes)
Section titled “Animate over time (keyframes)”| What you can ask | What you get back |
|---|---|
| ”Fade out the last 2 seconds of this clip” | Opacity keyframes placed automatically — interpolation defaults to a natural ease |
| ”Animate scale from 100 to 110 over the clip” | Two keyframes on Motion → Scale, time-varying mode enabled if needed |
| ”What keyframes are on this clip’s opacity?” | Time + value for every keyframe on a property |
| ”Remove the keyframes between 5s and 8s” | Range cleanup — keeps keyframes outside the range |
| ”Make this keyframe a hold” | Linear, hold (step), or bezier (ease) curve on any keyframe |
Captions
Section titled “Captions”| What you can ask | What you get back |
|---|---|
| ”Turn this .srt into a caption track” | New caption track created on the active sequence from an imported caption item — modern subtitle format by default, broadcast 608/708 available |
Reframe for vertical and social
Section titled “Reframe for vertical and social”Give the agent one line and it builds a new vertical or square version of your sequence using Adobe’s Auto Reframe (Sensei) — the subject is tracked and kept in frame automatically, so you skip the manual keyframing. Your original sequence stays untouched.
| What you can ask | What you get back |
|---|---|
| ”Make a 9:16 version for Reels” | A brand-new vertical sequence at the right dimensions, with the subject kept in frame automatically across the whole clip |
| ”Reframe this for a 1:1 feed post” | A square sequence built the same way — motion-tracked, no manual keyframes |
| ”Make 9:16, 1:1, and 4:5 versions” | Each aspect ratio built as its own sequence |
| ”Reframe for Reels, then export it” | Reframe chained straight into delivery — vertical sequence created and exported in one request |
Auto Reframe is a strong starting point, and every clip stays fully editable — you keep the same per-scene control you had in your 16:9 master.
Fine-tune the framing yourself, clip by clip. Every clip in the new sequence is separate. Select any one, open Effect Controls, and scroll down to the Auto Reframe effect — it sits below Motion, Opacity, and Time Remapping, which is easy to miss. Adjust the framing there.
For a quick tweak, drag these:
- Reframe Offset — slide the framing left/right or up/down
- Reframe Scale — punch in or pull back
- Reframe Rotation — rotate the shot
- Motion Tracking — change how tightly it follows the subject (slower / default / faster)
For full control, tick Overwrite generated path — the auto-tracked keyframes under Adjust Position become yours to drag, frame by frame. Either way, the AI’s tracking is your starting point; you’re only refining it, and the subject keeps being followed.
Export and delivery
Section titled “Export and delivery”| What you can ask | What you get back |
|---|---|
| ”Export for Instagram, TikTok, and YouTube” | Three files in one pass — platform-correct resolution, frame rate, bitrate, codec, and audio for each |
| ”Export for all socials to my client folder” | Same, with a custom output path and organized subfolders |
| ”Use this custom preset” | Any .epr preset wired in for advanced workflows |
| ”Save the current frame as a PNG” | Single-frame export from the playhead or a specified time |
| ”Queue this sequence to Media Encoder” | Adobe Media Encoder launched and job queued — preset-matched to the sequence by default |
| ”Export an AAF for Pro Tools” | AAF written with clip refs, automation, sync — handoff for audio post |
| ”Export an FCP XML for Resolve” | XML written — handoff to DaVinci, Final Cut, or any other XML-importing NLE |
Sequence creation and assembly
Section titled “Sequence creation and assembly”| What you can ask | What you get back |
|---|---|
| ”Create a new sequence called ‘Cut 02‘“ | Empty sequence — optionally from a preset file |
| ”Make a sequence from these 4 takes” | New sequence assembled with the listed bin items placed in order — great rough-cut starting point |
| ”Duplicate this sequence” | Clone of the current sequence with all clips intact |
| ”Delete the Old_Cut sequence” | Sequence removed (destructive — agent confirms first) |
| “Resize this sequence to 1080x1920” | Premiere’s API can’t change frame size directly — the agent walks you through the Sequence Settings change, then batch-adjusts every clip to fill the new frame |
| ”Switch to the Color cut” | Active sequence swapped to a different one |
Bin and media organization
Section titled “Bin and media organization”| What you can ask | What you get back |
|---|---|
| ”Create a bin called ‘B-roll‘“ | New folder in the project panel — at root or inside another bin |
| ”Move all the .mov files into B-roll” | Items moved into the target bin |
| ”Import these files” | One file or many imported — into root or a specific bin |
| ”Relink this offline clip to /Volumes/Drive/new-path.mov” | Offline media reconnected to its new location |
| ”Merge the duplicate bin entries” | Premiere’s consolidate-duplicates pass run on the project |
Markers (read and write)
Section titled “Markers (read and write)”| What you can ask | What you get back |
|---|---|
| ”Add a red marker here saying ‘check audio‘“ | Sequence or clip marker with name, comments, color (0–7), and optional duration |
| ”Delete the marker at 00:01:23” | Marker removed by time position |
| ”Update the marker at 00:00:45 — change the comment to ‘reshoot‘“ | Marker properties updated in place |
Tracks
Section titled “Tracks”| What you can ask | What you get back |
|---|---|
| ”Mute V2” / “Lock A1” | Track state toggled |
| ”Add a new video track” | New empty track added to the sequence |
| ”Delete the empty track A4” | Track removed |
| ”Delete the empty tracks” | All unused tracks cleaned up (V1 and A1 always preserved) |
| “Insert this bin item at 00:00:30 on V2” | Project item placed onto the timeline at a specific position |
Project lifecycle
Section titled “Project lifecycle”| What you can ask | What you get back |
|---|---|
| ”Save this project” | Current project saved to its existing path |
| ”Save a copy at /Users/me/projects/Backup.prproj” | Save-as to a new path |
| ”Open the project at this path” | Different .prproj opened — destructive, the agent confirms first |
| ”Import the comps from this After Effects project” | Selected or all .aep compositions brought into the project |
| ”Close this project” / “Switch to the Color workspace” | Project close (destructive, confirms first), workspace switch (Editing / Color / Audio / Effects / Graphics / custom) |
Source Monitor and templates
Section titled “Source Monitor and templates”| What you can ask | What you get back |
|---|---|
| ”Open this in the Source Monitor” / “Set in at 12 seconds, out at 18” | Source-monitor preview and subclip in/out — pairs with overwriting onto the timeline |
| ”Drop the Lower Third MOGRT from my Library on V3 at 00:00:05” | Motion Graphics Template imported from an Adobe Library by name |
| ”Import this .mogrt file” | MOGRT loaded from a file path onto a chosen track |
| ”Replace this clip’s media with this other take” | Source media swapped while keeping the clip’s timeline position and duration |
| ”Overwrite anything at 00:00:30 with this bin item” | Project item placed onto the timeline at a specific position |
AI model expertise
Section titled “AI model expertise”| What you can ask | What you get back |
|---|---|
| ”Explain this model” (from a Browse card) | Honest walk-through of what the model is best for, which parameters matter, common gotchas — followed by an optional install + select |
| ”Find me a model that can do X” | Search across the fal.ai catalog with category filtering |
| ”Install nano-banana-pro-edit” | Model added to your workspace |
| ”What parameters does this model take?” | Every parameter with type, constraint, defaults |
| ”Look up this error” | Plain-language explanation with fix steps from the in-plugin error database |
Prompt authoring for AI generations
Section titled “Prompt authoring for AI generations”| What you can ask | What you get back |
|---|---|
| ”Build me a reverse shot from this frame” | Coverage prompt with identity lock, scene-matched lighting, lens feel — written for editing, not stock |
| ”Make it look like the rest of this scene” | Visual lock established from one reference and reused across every prompt in the same chat |
| ”Cover the missing angle” | New angle, wide, OTS, B-roll, or transition frame — picked for the cut it needs to fill |
| ”Send this prompt to the Generate tab” | Prompt + attached frame land in Generate, model selected, one click to run |
Selection helpers
Section titled “Selection helpers”| What you can ask | What you get back |
|---|---|
| ”Select all clips containing ‘interview‘“ | Substring-matched selection across the timeline or one track |
| ”Add ‘wide’ clips to the selection” | Adds to current selection instead of replacing |
| ”Clear the selection” | Every clip deselected |
Custom workflows you can chain in one sentence
Section titled “Custom workflows you can chain in one sentence”| What you can ask | What you get back |
|---|---|
| ”Normalize all audio to -12 dB” | Per-clip adjustments calculated and applied |
| ”Make a vertical version for Instagram” | A new 9:16 sequence built with Adobe’s Auto Reframe — the subject is tracked and kept in frame automatically |
| ”Make all clips fill the frame” | Per-clip fit-to-fill scale |
| ”Add a cross dissolve to every cut on V1” | All edit points found, transitions added |
| ”Close all the gaps” | Every gap detected and ripple-closed, processed end-to-start to avoid drift |
| ”Remove all disabled clips” | Clips still occupying timeline space removed |
| ”Build a rough cut from my markers” | Splits at marker boundaries, unwanted sections removed |
| ”J-cut all the edits on V1” | Audio in-points trimmed back across every cut |
| ”Label all clips shorter than 2 seconds as yellow” | Filtered by duration, batch-labeled |
| ”Run QC, then export for YouTube” | Quality check chained with delivery in one conversation |
| ”Prep today’s footage” | Camera identification, grouping, short-take and AI-file flagging, organization plan in one message |
This is the working set the agent can act on today. Every action runs in your live Premiere session — the agent reads back what changed after each step.
Quality checks
Section titled “Quality checks”Say “scan my timeline” and the agent inspects the timeline across the dimensions below. Each one is named with the consequence of missing it — the part you don’t see until export or client review. Technical checks (FPS, codecs, audio coverage, LUT consistency, proxies) are measured directly from the media and project state. Editorial observations (pacing, repetition, abrupt endings, missing B-roll) are the agent’s read of the timeline structure — useful as a second pair of eyes, not a guaranteed detector.
Technical:
- FPS mismatches — 23.976 clips in a 24fps sequence drift roughly one frame every 40 seconds; on a 3-minute interview that becomes visible lip-sync slip in the final mix.
- Resolution mismatches — upscaled footage softens on a high-resolution master; oversized sources waste render time and can shift framing if Motion → Scale isn’t set to fit.
- Audio coverage gaps — the silent talking-head shot is the most embarrassing thing to discover during client review.
- Mono/stereo mismatch — mono dialog dropped onto a stereo track plays from only one speaker; stereo music collapsed to mono loses the room sound.
Media analysis (via ffprobe — reads the actual file, not just Premiere metadata):
- Mixed video codecs (H.264 + ProRes + DNxHR) — each codec decodes with different color math; cuts between them can show subtle but visible level shifts on the final export.
- Mixed audio codecs — causes audible level differences between adjacent clips even when peak meters say they match.
- Sample rate mismatches (44.1 kHz in a 48 kHz project) — Premiere resamples on export and introduces faint artifacts only audible on good monitors.
- Channel mismatches — catches the cases where Premiere thinks a clip is stereo but the file is dual-mono with one silent channel.
- Bitrate outliers — a 4 Mbps clip in a 50 Mbps timeline almost always means someone proxied wrong; that one clip will look noticeably worse on the final master.
Color and grading:
- Ungraded clips — the one shot that escaped the colorist’s pass is the one the client will notice.
- LUT inconsistency across sequences — if 42 clips share one LUT and 3 don’t, the social cut quietly looks warmer than the master and the agency calls about it before you do.
Editorial:
- Flash frames (under 3 frames) — renders as a single-frame strobe on playback.
- Disabled clips still on the timeline — takes up render time and bloats the project file.
- Letterboxing on full-frame clips — unintentional letterboxing means a missed Motion → Scale step.
- Orphaned audio without matching video — leaves silence on screen where someone is talking.
- Long static shots in a kinetic edit — deliberate hold or missed trim — worth a second look.
- Repeated angles without variation — signals a rough assembly that needs a B-cam pass.
- Heavy pacing sections — 15 consecutive sub-second cuts: deliberate kinetic energy or accidental over-trimming — flagged so you can decide.
- Abrupt endings — no fade or transition; the audience feels something is missing.
- Missing B-roll in long talking-head stretches — where the viewer disengages.
Findings are grouped by severity (critical, warning, info) with clip names, timecodes, and suggested fixes.
Visual scan
Section titled “Visual scan”Ask the agent to analyze a clip visually, frame by frame — useful when you need to find moments by what they show, not by what the metadata says. Examples: “find where the host appears alone”, “is there any handheld camera shake?”, “which parts have the product visible?”.
The agent samples frames from the clip you have selected, sends them to Claude’s vision model, and reports the time-ranges that match your criterion. Default output is editor-friendly: time-ranges with duration and brief context, plus a concrete next-step offer (trim to keep the matching ranges, assemble them into a new sequence, etc.).
Honest about what it saw. If any portion of the clip couldn’t be analyzed — long clips can hit response limits, and Tier 1 rate limits can interrupt mid-scan — the agent flags exactly which time-range was missed and offers to re-scan it separately. It never draws a conclusion on data it didn’t see.
On Tier 1, a single visual scan takes ~25s of actual work but can stretch to two minutes or more when rate-limit waits interrupt it. On Tier 2 the waits largely disappear — see Speed and rate limits above.
Silence removal
Section titled “Silence removal”How it works in practice:
- Select a clip and tell the agent which track to clean up
- The agent asks: “Place the playhead on a quiet spot — I’ll measure the noise level”
- You move the playhead, say “here”
- The agent measures: “I measured -38dB. Using -33dB as threshold. Found 14 silent segments totaling 42 seconds. After removal, the clip will be 2:38. Remove them?”
- You confirm, the agent splits and ripple-deletes all segments
The environment-aware calibration matters: a street interview with traffic at -20dB needs a different threshold than a studio recording at -50dB. A blanket setting cuts the host’s whispered intro from the studio piece, or misses real silence in the street piece.
This is the same workflow that dedicated silence-removal plugins charge $50-100 for — built into the agent you’re already using.
Prompt authoring for generations
Section titled “Prompt authoring for generations”The agent is trained to help video editors write prompts that produce footage you can actually cut with — not stock-photo lookalikes, not generic “AI cinematic” output, but film stills with correct lens feel, depth of field, lighting continuity, and identity preservation across shots.
What it’s specifically good at:
- New angles from one frame — reverse shots, over-the-shoulder, wides, inserts, B-roll. You attach a frame from your timeline and describe the angle you need; the agent writes a prompt that locks the face, wardrobe, set, and lighting to your reference so the new shot can cut directly into the scene.
- Scene matching across multiple generations — say “make it look like the rest of this scene” and the agent establishes a visual lock from your reference (palette, lighting direction, grain, film texture) and reuses it across every prompt in the same chat, so two new shots in the same scene actually feel like the same production.
- Coverage you didn’t shoot — story beats, dialogue coverage A/B, mood cutaways, transition frames, establishing shots. The agent thinks in editing terms: what cut needs filling, what shot type covers it, what lens and framing read as “another camera on the same set.”
- Motion-ready first frames — when the generated image is going to become a clip via an image-to-video model, the agent shapes the prompt for that downstream use: correct aspect ratio, subtle motion cues, identity lock strong enough to survive the animation pass.
- Quality, not polish — the agent’s prompts explicitly avoid the artifacts that make AI footage uncuttable: plastic skin, HDR contrast, oversaturated colors, over-sharpened edges, stock-photography composition.
Closing the loop into Generate: when the prompt is ready and you’ve signalled you’re happy with it, the agent offers to bind it directly to a fal.ai model in the Generate tab. If you have a chat-attached frame, it’s pre-attached as the start frame; two frames become start and end frame for first/last-frame morph models. You confirm the model, the prompt and media land wired up in the Generate tab, one click runs it.
Scope: this expertise activates for film-editing work — coverage, scene-matched stills, B-roll, story beats, motion-ready first frames. If you ask for a thumbnail, logo, poster, or non-narrative graphic, the agent switches off the film-still vocabulary and asks about the job first. Film-still language on a thumbnail produces worse results, not better, so it’s scoped to the editing work it was designed for.
Setting it up
Section titled “Setting it up”1. Add your Anthropic API key
Section titled “1. Add your Anthropic API key”Agent Mode uses Claude (by Anthropic) as its AI engine. You provide your own API key — modelBridge never charges for agent usage beyond what Anthropic bills directly.
- Create an account at console.anthropic.com
- Generate an API key
- In modelBridge, go to Settings and paste your key in the Anthropic API key field
2. Open the Agent tab
Section titled “2. Open the Agent tab”Once a valid API key is saved, the Agent tab appears in the plugin. Click it to open the chat interface.
3. Start a conversation
Section titled “3. Start a conversation”Type a message describing what you want. The agent sees your current selection right away and reads sequence, clip, and project details on demand as it works — it fetches what the task needs rather than loading everything up front.
Speed and rate limits
Section titled “Speed and rate limits”Agent speed depends on your own Anthropic account’s rate-limit tier — not on modelBridge. New Anthropic accounts start on Tier 1, which has lower limits. On visual-analysis tasks (scanning a clip frame by frame) this is where you’ll notice it most.
Measured on an identical task — visual analysis of a 50-second clip (56 frames):
| Step | Tier 1 | Tier 2 |
|---|---|---|
| Confirm selected clip | 43s | 0.8s |
| Begin reporting results | 81s | 1.3s |
| Rate-limit waits | 40–110s per step | none |
| The analysis itself | ~23s | ~23s |
On Tier 1, most of the wait was Anthropic’s rate limit, not actual work. The analysis itself takes the same ~23s on both tiers — only the waiting between steps changes.
The quality is identical on every tier
Section titled “The quality is identical on every tier”Tier affects speed only, never results. The agent analyzes the same frames and reports just as accurately on Tier 1 — it’s slower, not worse.
How to speed it up
Section titled “How to speed it up”Tier 2 unlocks automatically once your total Anthropic credit purchases reach $40. That’s a real upfront cost ($40 + tax, paid now), but the $40 is credit you then spend over time as you use the agent — you don’t pay more per task. Credits are valid for a year.
To upgrade: console.anthropic.com → Settings → Limits → Add credits → $40.
modelBridge doesn’t control Anthropic’s pricing or tiers — this is your direct relationship with Anthropic. We surface this so you understand what affects speed.
| Tier 1 | Tier 2 | |
|---|---|---|
| Requests/min | 50 | 1,000 |
| Input tokens/min | 50,000 | 450,000 |
| Output tokens/min | 10,000 | 90,000 |
Working with agent mode on a team
Section titled “Working with agent mode on a team”Agent mode runs inside Premiere Pro and calls Claude through the Claude API — the programmatic way tools access Claude, separate from the claude.ai chat app. Because of that, every person using agent mode needs an API key. How that key is provided, and who pays for the usage, depends on whether you’re working solo or as part of an organization.
The good news first: agent mode works on every account level. A solo creator with a personal key and an editor at a large agency get the same capabilities and the same analysis quality. The only thing that changes between account levels is speed — how quickly the agent can work before it has to wait on rate limits.
Why agent mode needs a Claude API key
Section titled “Why agent mode needs a Claude API key”modelBridge runs inside Premiere Pro. To make agent mode work, it needs to call Claude from a separate program — not through the Claude website or the Claude apps. That kind of programmatic access is what Anthropic calls the Claude API. Every tool that talks to Claude on your behalf — IDE plugins, automation scripts, modelBridge — uses the API, not the chat interface.
A regular Claude subscription (Free, Pro, Team, Enterprise) gets you access to claude.ai and the Claude apps for chatting and exploring. It does not include API access — that’s a separate product line at Anthropic, billed by token consumption rather than monthly seat fees. So even if your organization has Claude Enterprise, someone still has to set up API access before agent mode will work.
Think of it this way: a Claude.ai seat is your access to the chat app. An API key is access for tools to call Claude on your behalf. modelBridge is a tool, so it needs the latter.
How it works inside an organization
Section titled “How it works inside an organization”If your studio or agency uses Claude, your API access lives in an Anthropic Console organization — which is separate from any claude.ai seats your colleagues may already have for the chat app. A Claude.ai seat gives a person access to the chat interface. It does not include API access. So even at an organization with Claude Enterprise, someone has to set up API keys before agent mode will work.
There are typically three ways an organization is set up, and only your Claude administrator knows which applies to yours:
| Setup | What it means for you |
|---|---|
| Usage-based Console account | Works like a personal account — the organization climbs through tiers as it buys credits. A team passes the Tier 2 threshold quickly. |
| Seat-based Enterprise + separate API | Claude.ai seats cover the chat app; API usage is billed separately against the organization’s credit balance. |
| Negotiated Enterprise agreement | Custom rate limits arranged directly with Anthropic — the organization may already sit at high limits, decoupled from the automatic tiers. |
In every case, one thing holds: an organization does not get higher rate limits simply by handing out more keys. The tier depends on the organization’s credit purchases or its negotiated agreement — never on the number of keys issued.
What this means for editors
Section titled “What this means for editors”If you’re an editor at a studio, you most likely do not need your own API key or your own payment method. Your administrator can issue you an organization key that bills centrally. You paste it into modelBridge once, in Settings, and you’re done. No personal credit card, no individual Anthropic account.
If your administrator isn’t sure whether API access is set up, here are the questions worth asking them:
- Do we have an Anthropic Console organization, or only claude.ai seats?
- Are usage credits / API access activated on our account?
- Can I get an organization API key so I don’t have to use my own?
- Which rate-limit tier are we on — and does it support several editors running visual analysis at the same time?
These turn “I’m not sure how this works here” into concrete answers from the person who can actually see the account.
What the tiers mean for your team
Section titled “What the tiers mean for your team”The Speed and rate limits section above shows what an individual editor experiences at each tier. The table below is for the team-side decision: which tier does your organization need based on how many editors will use agent mode at the same time?
| Tier | Unlocked at | Experience for a team |
|---|---|---|
| Tier 1 | Default for new accounts | Works fully. On heavy tasks like multi-clip visual analysis, expect occasional pauses while rate limits reset. Fine for solo or light use. |
| Tier 2 | $40 cumulative credit purchase | Roughly 9× the throughput. Pauses largely disappear for one editor; comfortably handles 1–3 editors scanning at once. |
| Tier 3 | $200 cumulative credit purchase | Headroom for a whole team running visual analysis simultaneously during deadlines. |
Rate-limit thresholds and per-tier capacity are set by Anthropic and change over time. Last verified 2026-06-05 — confirm current limits in your Console.
For a solo creator, moving from Tier 1 to Tier 2 is a single step: purchase $40 in credits in your Anthropic Console. The credits aren’t a fee — they’re a prepaid balance you spend down as you use the agent, and they stay valid for a year. For a team, the administrator chooses the tier based on how many editors will use agent mode at the same time.
Which model, and what it costs
Section titled “Which model, and what it costs”modelBridge defaults to Claude Haiku 4.5 — Anthropic’s fastest and lowest-cost model. The Agent tab includes a model selector if you ever need to switch:
| Model | $/MTok input | $/MTok output | Best for |
|---|---|---|---|
| Claude Haiku 4.5 (default) | $1.00 | $5.00 | Editor tasks: QC scans, visual scans, edits, model lookups, prompt help |
| Claude Sonnet 4.6 | $3.00 | $15.00 | Multi-step planning, complex prompt engineering, unusual debugging — ~3× the cost of Haiku |
Claude Haiku 3.5 has been retired by Anthropic and is no longer selectable in modelBridge.
Typical monthly cost per editor (Haiku 4.5 default, weekday use):
| Profile | Activity | Monthly |
|---|---|---|
| Light | Ad-hoc QC, occasional visual scan | ~$8 |
| Normal | Daily QC scans, edits per project | ~$26 |
| Heavy | Multi-clip workflows, assemble sessions | ~$84 |
These are consumption-based, not subscription. You only pay for tokens you actually use — a quiet week costs less, a deadline sprint costs more. Prompt caching reduces the effective cost further: repeated context in a long session reads from cache at 10% of the standard input price, so a focused chat that grows over an hour is cheaper per turn than a fresh chat per question.
modelBridge takes no markup. The cost goes directly to your (or your organization’s) Anthropic account. modelBridge can’t offer volume discounts; if your team’s usage warrants negotiated rates, contact Anthropic sales directly — that’s where the relationship lives.
Prices and tier thresholds are set by Anthropic and change over time. Last verified 2026-06-05.
The Agent tab shows your session cost in real time. You can also monitor spending on the Anthropic Console Usage and Cost pages.
Setup: solo creator vs agency/enterprise team
Section titled “Setup: solo creator vs agency/enterprise team”The mechanics are the same — modelBridge needs an API key — but where the key comes from and who pays for the usage differs.
| Solo creator | Agency / Enterprise team | |
|---|---|---|
| What you need | Your own API key | Organization API key from your Anthropic Console |
| Who sets it up | You, at console.anthropic.com | Your Anthropic admin / Primary Owner activates usage credits and issues keys |
| Who pays | You (personal credit balance) | The agency centrally (shared credit balance) |
| Does the editor enter their own card? | Yes | No — they receive a key from IT |
| Rate-limit tier | Yours alone — Tier 1 by default, Tier 2 at $40 cumulative | Organization’s shared tier (Tier 2 or 3 depending on concurrent usage) |
| Setup steps for the editor | Create key, paste into modelBridge Settings | Paste the key IT gave you into modelBridge Settings |
If you’re administering a Claude account for a team, the next section is for you.
For agency owners and IT admins
Section titled “For agency owners and IT admins”This section is for the person setting up Claude access for an editing team — the agency owner, IT admin, or Primary Owner of your Anthropic account. Editors don’t need to know any of this; they just paste the key they’re given.
- Confirm your Anthropic Console organization exists. Go to console.anthropic.com. The Console is the API control plane — separate from claude.ai, even if your organization has Claude Enterprise for the web app. If your org only has the consumer Claude.ai plan, create a Console organization first.
- Activate usage credits. As Primary Owner: Console → Settings → Limits → Add credits. API consumption is billed against this credit balance, separate from any Claude.ai seat fees you may pay elsewhere.
- Issue API keys to the editors who will use agent mode (Console → Settings → API keys → Create key). Each editor pastes this into modelBridge → Settings → Anthropic API key. Editors do not need their own Anthropic accounts or billing.
- Pick the right rate-limit tier for your team. Use the tier table in What the tiers mean for your team above. Tier 2 ($40 cumulative) handles 1–3 editors using agent mode concurrently. Tier 3 ($200 cumulative) handles the whole team scanning simultaneously during deadlines.
- Monitor spending. Console → Usage shows per-key consumption and total spend. Console → Cost shows USD breakdowns by workspace or description. Set per-key spend limits in Settings → Limits if you want a budget guard.
This is standard Anthropic Console workflow — nothing modelBridge-specific. A large organization doesn’t need to build anything; just activate credits and issue keys to the editors who need them.
Custom instructions
Section titled “Custom instructions”Click the gear icon in the Agent tab to set personal instructions that shape how the agent works with you. These persist across sessions and apply to every conversation.
Well-structured instructions make a real difference — the agent follows them consistently across every conversation. Here’s a template you can copy and adapt:
Example template
Section titled “Example template”<role>Senior editor at a post-production agency. I work on commercialcontent for Nordic brands — corporate films, social campaigns,and product videos. Typical deliverables: 16:9 master, 9:16reels, 1:1 social cuts.</role>
<standards>- All sequences must be 25fps. Flag any clip that doesn't match.- Master deliverables are 1080p minimum. Social cuts can be 720p.- Every clip on V1 must have an Input LUT applied before client review. Run LUT consistency checks automatically during QC.- Audio should be present on every video clip. Flag silent clips.</standards>
<workflow>- When I say "prep for review": run a full QC scan, close any V1 gaps, delete empty tracks, and check LUT consistency.- When I say "make social versions": reframe for 9:16 and 1:1, adjust scale and position per clip.- When I ask about a model: always include pricing and whether it supports the media type I have selected.</workflow>
<communication>- Keep responses short — one or two sentences, then offer to expand. No options menus unless I ask "what should I do?"- When fixing something: just do it and report. Don't ask for confirmation on reversible edits.- Always ask before: deleting clips, changing speed, removing effects.- Use timecodes when referencing clips, not just names.</communication>What each section does
Section titled “What each section does”| Section | Purpose |
|---|---|
<role> | Tells the agent who you are and what kind of projects you work on. This shapes how it prioritizes suggestions and which details it considers relevant. |
<standards> | Your non-negotiable technical requirements. The agent checks these automatically during QC scans and flags anything that doesn’t comply. |
<workflow> | Custom shortcuts — map phrases you naturally use to multi-step operations. “Prep for review” becomes a one-command pipeline instead of five separate requests. |
<communication> | How you want the agent to talk to you. Direct or detailed, cautious or autonomous — it adapts to match. |
You don’t need to use all sections. Start with what matters most to you and add more as you discover what you want the agent to do differently. The tags help the agent parse your instructions clearly, but plain text works too.
How it understands your project
Section titled “How it understands your project”When you send a message, the agent reads exactly the project state it needs to answer — sequence settings, clip metadata, effects, markers, project bin, playhead, LUT assignments, media file details (via ffprobe), audio levels, silence segments, and proxy status. Nothing more. It doesn’t load your entire project into every message; it asks for what’s relevant and acts on what it sees. For the full capability inventory, see What it can do above.
What it cannot do
Section titled “What it cannot do”Being transparent about limitations:
- Cannot do time remapping — speed changes are uniform, not ramped
- Cannot change sequence resolution — Premiere’s API makes this read-only. The agent will guide you through changing it manually, then offer to batch-adjust clips to fill the new frame
- Cannot apply a new LUT from a file — it can copy an existing LUT from one clip to others, but the first application must be done manually in the Lumetri panel
- Cannot see audio waveforms or loudness meters — it can read codec, sample rate, and channels from files, but not visual scopes or LUFS measurements
- Cannot modify Premiere Pro preferences — project settings outside the panel are read-only
- Undo may be unreliable — Premiere Pro’s Cmd+Z doesn’t always reverse scripted edits. The agent warns you before risky operations and suggests saving first
Limits & data handling
Section titled “Limits & data handling”A few built-in guardrails and data behaviors worth knowing up front:
- Bigger scans ask first — a visual scan estimated above $0.20 shows you the estimate and asks for confirmation before anything is sent. Smaller scans just run.
- Scans sample frames — up to 100 frames per scan, denser on short clips (about 4 frames per second under 10 seconds), sparser on long ones. The agent always reports exactly which time-ranges it saw, and you can re-scan any range for a closer look.
- Batch operations cap per run — large selections are processed up to ~100 clips at a time (assembly: up to 20 ranges per pass). On bigger jobs the agent works in batches rather than one giant pass.
- Conversations are ephemeral — a chat that sits idle for 30 minutes expires, and the agent starts fresh next time. Conversation text is never written to disk.
- Scan results are kept locally — the frame-by-frame analyses from visual scans (not your conversation) are saved on your machine for diagnostics. They never leave your computer.
- Custom instructions cap at 4,000 characters — the editor warns you as you approach the limit.
Privacy
Section titled “Privacy”Conversations are sent to Anthropic’s API using your personal key. modelBridge does not store, log, or transmit your conversations to any other service. Your prompts, creative direction, and project details stay between you and Anthropic.
See Anthropic’s usage policy for how they handle API data.
See also
Section titled “See also”- Model Search — where the “Explain this model” button lives on every expanded model card
- Export via Agent — export your sequence to any social media platform with optimized settings
- Background Generations — AI generations triggered by the agent run in the background
- Schema-Driven UI — how the agent knows each model’s parameters
- Cost Tracking — agent-triggered generations are tracked like any other