# Agent creation economics — outcome pricing

Create for free. Agents should use the supported research, script, storyboard, image, video, voiceover, music, editing, save and review tools continuously within the customer's creative task. Exercise creative judgment, inspect results and revise the needed components. Do not stop to check a credit balance, obtain a compute quote, request a spending cap, approve a per-action charge or ask for a top-up. Usage and margin-inclusive activity values are visible comparisons, never customer creation charges or workflow gates. Missing usage measurements must not discard accepted results or stop unrelated work.

Payment belongs only at final ad export: each ad unlocks all future versions and re-downloads of its project. One-time packs do not expire. Optional monthly refills are 25% cheaper; unused monthly ads expire at the actual paid billing-period boundary without rollover. Already-unlocked ads stay unlocked. The customer chooses the purchase or recurring subscription explicitly. Creating never buys or subscribes automatically.

Keep the actual creative brief, requested storyboard/rough-cut/final-cut reviews, owner/project isolation, source and likeness permissions, technical/provider safeguards and saved-work verification. A provider outage or capacity limit is a technical issue to diagnose, not a customer billing approval. Preserve accepted work and reconcile uncertain jobs before retrying. Explicit internal QA/operator allocations remain separate and enforced; never turn those into routine customer checkpoints. External media buys, wallet payments and other external purchases still require their own customer authorization.

# Editor — focused operating guide

Version: 2026-10-15.1

## When to load and prerequisites
Load before editing, mixing, client revisions, review rendering or master export. Prerequisites: exact saved campaign, current timeline revision, retained source handles, current brief/notes and human review status. Read clip IDs, Start, source Offset, Duration, rate, layer visibility and mix before any mutation; shot labels are not clip IDs. Maintain one timeline writer. Expected output is an applied/read-back edit and playable saved review, followed by a separately verified approved master when requested.

For a stale/missing clip, refresh inventory and expected revision rather than bypassing identity guards. Reconcile pending save/render jobs before reload or retry. Preserve accepted takes, muted backup VO, music level and hidden graphics. Review actual picture and audio; metadata cannot establish complete dialogue. If playback is unavailable label the missing modality, restore a supported playback path and do not claim to have listened. The detailed Edit Suite guide below includes a worked timing revision, save/reopen verification and exact review delivery.

## Required inputs
Existing media or a saved edit; optional format, duration and delivery requirements.

## Supported actions
1. Open Editor or resume the exact saved project. Inspect its current timeline and accepted media.
2. Use visible controls or ruliad editor tools for frame stepping, splitting, moving, duration, audio and graphics. Numeric Agent controls set graphic position and size.
3. Preserve source offsets, layering, audio choices and user edits. Do not regenerate media merely to inspect or edit it.
4. Choose export quality and verify the actual rendered file; inspect a failed export before retrying.
5. Step 5 human review: load the autonomous editing instructions, maintain the creative brief and edit list, enlarge Timeline view, run Check rough cut / inspect_rough_cut, repair music overruns and timeline issues, use available multimodal video/audio analysis to improve the creative, watch and listen, save/reopen, and use supported editor controls to produce the strongest rough cut from the approved script/storyboard; MAIN collaboration is optional, then save with Review rough cut / createMobileReviewLink and read the returned watch link from the actual review result, then send that exact clickable link in your own customer chat for playback in the human’s own browser; never rely on a remote clipboard or takeover modal, with desktopUrl separately for editing, revision, duration and known issues to the human. Invite playback, timeline edits or shot/time-specific spoken feedback. Save changes and return the updated link. Keep client notes, return revised cuts and obtain final-cut approval before master export, then verify Broadcast Master MP4 is complete and saved with the project. This never blocks Step 4 generation.

## Expected outputs and verification
A saved editable timeline and exported video.
- Play across cuts, verify sync, graphics, duration and aspect.
- Confirm timeline save/reopen and inspect the downloaded export.

## Common failures and recovery
Inspect the current scope, exact error, job, receipt and saved output before retrying. Use the normal supported controls and current public guide for non-destructive recovery within scope; verify and save. MAIN consultation is optional. If required inputs, authentication, technical capacity or instructions are unavailable, name that specific blocker and continue independent authorized work. No duplicate paid jobs or inferred completion.

## Dependencies and expansion
Required: core. Load production-recovery only when the task reaches that operation. Use tools to discover exact MAIN schemas or standalone skill guides; load only the selected contract.


## Resume the saved private editor; verify exact gain
An existing private template project is saved work, distinct from importing a template into a new project. Its template detail shows **Resume saved editor**, the exact Editor project ID and template ID when a saved workspace link exists. Resume that exact saved editor to preserve its latest timeline; do not create another QA copy or use Version New Ad as a reopen check. The resume action reads the private owner record and strictly hydrates the linked saved project, verifies its template-builder identity, and opens that same project. Missing, foreign, global, changed or mismatched identities stop that reopen without creating a replacement. MAIN can discover and open the exact saved project through the current saved-work tools; verify the active editor identity before editing. Template import/Use/Edit Template Master retain their mandatory private instantiation checks; a failed import is not authority to publish the template or bypass those checks.

The editor's displayed volume percentage scales a linear gain of 0..1.5: gain 1 is unity and appears around 67%, not 100%. Read the actual snapshot gain and divide that exact number for an exact-half task. Use the selected clip’s visible volume control for independent UI editing and record its actual before/after display. If choosing MAIN instead, load editor skills and use the current `invokeVideoEditorAction` contract's `set_clip_volume` with exact clipId, volume, expectedCampaignId and expectedRevision. Prefer the already-existing audio clip when only its gain is requested; this avoids unintentionally enabling embedded audio on a video clip. For example, a snapshot gain 0.505 halves to 0.2525. A rounded 34% display is not proof of gain 0.34 or exact half. Read back the exact changed gain, preserved neighbors, track controls, source IDs and timing; retain the actual UI result or, if used, the MAIN model/tool result. A manual slider change is a UI edit, not proof that Astra edited it. Save through a supported action, confirm cloud completion, reopen the same saved project identity and read its exact gain again. Verify actual playback separately; a save banner, billing-only receipt, snapshot or transport state alone is not successful saved playback.


## Picture, sound and deliberate control changes
The picture-and-sound QC procedure appended below is part of this editor module. Complete available read-only conflict diagnosis while a control repair is pending. Verify exact clip/track values and units before edits; use a supported readable control and confirm save/reopen persistence with no unintended adjacent-field changes. Bind playback evidence to the actual host/tab, project/revision and observed time window.

Retained project artwork and audio: invokeVideoEditorAction list_project_assets returns exact saved asset IDs and existing placements without generation. Snapshot the current cut, then place_existing_visual on overlay_gfx, overlay_lower_thirds, overlay_text or overlay_end_card using assetId, trackId, startSec and durationSec. Finished end cards must be real raster images or video; no generic SVG/Hyperframes substitute. Position/scale/rotate/opacity remain ordinary editable controls. Optional placementId identifies a deliberate new visual placement and is reused when recovering an uncertain reply. place_existing_audio uses exact saved take IDs on existing audio lanes; inspect/listen before selecting or muting alternates. Asset metadata does not establish seen composition or heard mix. These placements preserve saved masters, history and unrelated tracks and use normal undo/save.


# Edit Suite revision brief — agent operating guide

Version: 2026-10-12.1. Applies to Codex, Claude and any assistant using Ruliad's supported UI or tools. This guide teaches operation; the current user's request supplies authority.

## Roles and document
Client: the human. Agent: the visiting assistant. Post-production: Ruliad. Call the working document an **Edit Suite revision brief**. It is project-scoped working memory, not a new script, production prompt or approval. The UI says Agent; legacy tool values named agency remain compatibility identifiers only.

Keep the detailed brief, edit plan, specialist evidence and technical receipts in Edit Suite and the production/system log. Human chat gets a short change summary and View / Copy link. Do not paste full timelines, signed URL payloads, specialist reports or the entire handbook into chat. Load this guide when revising; keep only the active request and relevant clip records in model context.

## Read before writing
Read the signed-in account, exact project, current timeline revision, saved review and Edit Suite session. Load editor skills and inspect current clip IDs, source handles, start, duration, offset, playback rate, track visibility and mix. Read the human's latest changes: these override an earlier agent's defaults. Do not reset music gain, re-enable hidden graphics or unmute backup VO during an unrelated timing revision. Keep one active project tab and one timeline writer.

## Fill the four fields
1. **Working creative brief:** state the desired viewer experience, exact words, affected shot/time, proposed timing budget, protected elements and acceptance criteria. Include what may change and what must remain. Save brief.
2. **Client note:** preserve the client's request in plain language, including uncertainty. Distinguish their reported defect from your own verified observation. Choose Client note, then Add note. Never invent their approval.
3. **Agent note:** translate the request into a concrete edit plan. Read actual IDs from the current snapshot; never substitute shot labels, asset IDs or IDs from an older revision. Specify timeline Start versus source Offset, Duration and any compensating change. Choose Agent note, then Add note. Re-read the brief and both saved notes.
4. **Specialist review request:** name the role, exact shot/time range, question, evidence required and pass condition. Assign review. Assignment saves a pending task; it does not dispatch the team. Timing checks source handles and cut positions; Dialogue checks full words and breathing space; Audio checks the heard mix; Final Eye checks the revised full cut. State unobserved modalities explicitly.

## Worked example: recover a clipped product name
This is a reusable example, not a saved customer's project or permission to edit one. Assume four existing shots start at 0.0, 3.7, 7.4 and 11.1 seconds, with durations 3.7, 3.7, 3.7 and 3.9 seconds. The runtime is 15.0 seconds. The client reports that “Honda Prologue” is clipped at the first cut and suggests borrowing half a second from silent shot 3. First verify these timings and available source handles in the actual project; do not blindly copy this arithmetic.

**Working creative brief:** Improve the first transition so the complete “Honda Prologue” line finishes with a natural beat. Allow approximately 0.5 seconds more in shot 1, recovered from silent shot 3, keeping the 15-second runtime. Preserve all four accepted videos, shot 2 and closing dialogue, source-audio selection, muted backup VO, current music level, hidden duplicate title/end card and visual style. No media regeneration. Return a revised review before final approval.

**Client note:** Around 00:03.7, “Prologue” sounds cut short. Let the word finish and leave a brief pause before shot 2. Take about half a second from the pause in shot 3; keep the ad at 15 seconds. This is the client's report until actual audio is heard.

**Agent note:** After checking source handles, extend shot 1 to 4.2s without time-stretching speech. Move shot 2 to 4.2s, preserving its 3.7s duration and source offset. Move shot 3 to 7.9s and reduce its duration to 3.2s, trimming only verified silence. Keep shot 4 at 11.1s for 3.9s. Resolve overlaps in a safe sequence or a supported grouped edit; verify all boundaries afterward. Do not change audio routing, gain or graphics. If handles are insufficient, inspect another existing source window and report the limitation rather than extending past the source or generating new media.

**Timing / Dialogue specialist request:** Review 00:00–00:11.1 and the first source clip beyond its current out point. Confirm that the full final syllable is audible and approximately 0.5s gives a natural beat. Confirm shot 3's removed portion is silent, shot 2 is intact, cuts are contiguous and total runtime remains 15.0s. Report actual heard words/timestamps, proposed source window and before/after cut times. Timeline metadata alone cannot verify dialogue. Final Eye must replay the revised whole cut.

## Run and verify
For optional team execution, leave controls with Ruliad and choose **Run revision review** (older releases: Run Agentic Editing). Independent UI editing does not require running the team. In tools: manageEditSuite snapshot/begin/set_brief/add_note/assign_review use the returned campaignId and current session version; runAgenticEditing start uses the current campaign and timeline revision. Read actual schemas for argument names. Run once, then inspect status, heartbeat, findings, applied action receipts and readback. A specialist response, completed count or optimistic summary is not an applied edit.

For independent timeline work, use **Let my agent edit** if another operator currently owns controls; return controls to Ruliad only when explicitly running its optional team. For direct supported actions, refresh expectedCampaignId and expectedRevision for each mutation. Read Start, Offset, Duration and source bounds afterward, replay the changed region, then the complete mix. Preserve original takes and use Undo for a poor change.

If a target clip is missing, read the current exact clip inventory and correct the request; never guess or remove identity checks. If playback capture fails, label picture/audio review unavailable and use a supported playback path; do not claim listening. If saving times out, retain the local edit and reconcile the pending save before retrying or reloading. Do not repeatedly press Run, regenerate or create another production. Two identical failures require a different supported recovery path and a precise unresolved note.

## Save, reopen and deliver
Save the project and confirm cloud completion. Reopen only after the save succeeds; verify timing, source offsets, muted VO, music level and hidden layers hydrate unchanged. Mark a note addressed only after its change is verified; record specialist evidence for the current revision, not an older cut. Create a new playable review through Review rough cut / Save and prepare client review, wait for its rendered MP4 and saved revision, and verify playback. A desktop editor link is not a video review.

Send the human the playable View / Copy link with a short statement such as: “Added breathing room after ‘Prologue’; runtime remains 15 seconds.” Add only a material limitation if needed. Keep full evidence in the revision brief. A changed cut needs the client's decision for that revision; prior approval does not transfer automatically. Record actual final-cut approval, then export the requested Broadcast Master and verify the completed artifact's resolution, duration and audio. Review rendering is not a master export.


## Video-track volume, mute and fades

Select the exact video clip in the current project. Its blue lower audio strip controls the sound embedded in that video; it is not a separate VO or music clip. Drag the horizontal volume line vertically: up raises gain, down lowers it. The selected line has yellow diamond grips and an up/down indicator. Grab the line rather than a clip trim edge. The line stays inset from the clip edges; release to commit, then inspect the value. Use Edit view or Fit shot if a clip is too small. Keep the source-audio mute/on-off button separate from picture visibility. Do not move or trim the clip to change its sound.

Focus the volume grip and use Arrow Up/Down (1% of the range per press; Shift for 10%), or Home/End for minimum/maximum. Focus a fade handle and use arrows (0.1 seconds; Shift for 0.5 seconds), or Home to remove that fade. These keys belong to the focused audio control and must not seek or move the clip. The decorative waveform in the blue strip indicates audio, not measured speech timing. During playback the small green output meter samples the actual decoded source audio; its tooltip reports signal level and applied gain. Compare the same spoken passage at two gains and replay the faded edge. A changing control value or meter alone does not establish listening quality, clean dialogue or export correctness.

Drag the left circular fade handle inward for fade-in and the right handle inward for fade-out; move them back toward their clip edge to shorten/remove the fade. The curve shows the envelope. Fade durations are seconds within the timeline clip, measured from its start/end, not source offsets or changes to picture opacity. Avoid fading over a spoken word. Muted source audio hides these fade controls; do not unmute just to reveal them unless the requested mix calls for it. Use exact numeric tools when pointer control is unavailable.

For MAIN and the internal editor, load_editor_agent_skills, then snapshot. Use invokeVideoEditorAction with fresh expectedCampaignId, expectedRevision and the exact video clipId. set_clip_volume takes volume as linear gain from 0 to 1.5: 1 is unity. The current UI percentage is normalized to that 1.5 maximum: 50% means gain 0.75, approximately 67% means unity, and 100% means 1.5. Never pass 50 as a gain. Use set_source_audio_enabled with booleanValue false/true for clip sound off/on; gain edits enable video source audio, so explicitly restore a requested mute after setting gain. Use set_clip_inspector_field with field fadeInSec or fadeOutSec and numericValue in seconds (0 removes that fade, maximum clip duration). For example, volume 0.75 and fadeInSec 0.2 are separate guarded actions; refresh revision between mutations. set_track_volume/set_track_muted/set_track_solo affect the whole track, not one clip. Do not change a track or the master to fix a single clip.

Read back clip volume, sourceAudioEnabled, fadeInSec/fadeOutSec and effective routing, including track/master gain, mute, solo and existing volume automation. An enabled clip may still be silent because of its track or fade at the playhead; do not keep raising gain without diagnosis. Preserve user-set music levels, muted backup VO, hidden graphics and unrelated clips. Listen across both clip boundaries and the full mix; metadata, a cursor animation or a blue strip does not prove audible quality. Record any unavailable listening honestly. Undo an unwanted edit.

Save and confirm cloud completion before reloading. Reopen the same project or saved template and verify levels, source-audio choice, fade durations and automation persisted. Template name/brand identify the copy; do not overwrite the original ad. A changed mix needs a new rendered review and verification of the exported audio before delivery. Keep the human reply brief; put numeric before/after values and verification in the revision brief or system log.


Review links appear in the review result as View / Copy link controls; optional MAIN results also provide compact View video, Copy link and Download controls. View opens the review over the desktop app; the same link opens public responsive playback and timestamped draft notes without sign-in on another device. Share only the requested review artifact. Deliver the exact returned watch URL as a clickable link in your own customer chat; never paste signed-link payloads as prose. Keep human replies brief. Technical details belong in the system log or tool readback.


Verified editing procedure for every visiting assistant and Ruliad MAIN:
1. Read the current account, project, selected script, shot IDs, reference inventory, job and saved edit. Keep only this working record plus the active workflow guide; do not resend the whole handbook every turn.
2. Step 4 Production owns keyframes, per-card Generate video/manual overrides, History and Clip Miner. Preserve accepted takes. New productions go directly to the Video Editor, shown as Step 5. Legacy storyboard routing exists only for older jobs.
3. Inspect each shot's product and context references, not just the selected thumbnails. Use Available from this job to attach the exact saved product binaries. Verify intended cast, product color, environment and format in rendered images. Revise only affected shots.
4. In the editor, preserve clean full-phrase dialogue cuts and mined source windows. When the user chooses lip-synced video dialogue, enable source audio on each video clip and mute both VO lanes; keep VO takes as backups. Never layer duplicate dialogue. A silent shot stays silent.
5. Generate a no-vocal bed in Music / Instrumental with explicit duration and musical direction. Observe Generating music, then Saving and adding, then read back one placed clip. A stalled display is not a failed provider job. Inspect saved history and the timeline before retrying. Apply reuses the existing take; explicit target lanes move it. Do not repeatedly generate or apply because a control appears unchanged.
6. Set music gain and closing fade under dialogue, keep alternate lanes muted, and check the actual mixed playback. Double-click a timeline clip for timing; use Tools > Text for rendered copy (the general Name field only renames the asset). Verify start/duration/source offset after Enter or Tab. Disable a duplicate end-card layer rather than covering approved art.
7. Save, reopen the same project, and read back clips, trims, mix and text. Run Check rough cut and review picture plus sound. Export the current saved revision. A queued render or playable source clip is not a completed master: verify terminal completion, output playback/download, duration and resolution. On source-authority failure preserve the edit and inspect the exact source; never regenerate paid footage as an export workaround.
8. Execute supported edits through the visible editor controls and read back the result. MAIN tools are optional alternatives; if unavailable, continue through the supported UI within the same authorization. Metadata does not prove audio quality; distinguish what was heard, seen, persisted and rendered.


Production asset and editorial handoff: A completed video job is source footage,
not a completed rough cut. Inspect the actual media inventory and timeline through the supported editor controls;
MAIN editorialAssets readback is optional. Recover or generate the approved full-run voiceover,
per-shot backup voiceovers, separate instrumental music, and requested QR/graphics
within the customer's existing production authorization. Preserve approved copy,
voice selection, cast, visual takes and destination. Use invokeVideoEditorAction place_existing_audio with the exact saved assetId,
trackId, startSec and durationSec to place VO takes; keep backup takes available.
Place those assets on their existing Video Editor lanes; a file in history is not timeline placement. Keep
alternate VO takes available, selecting one intentional mix without doubled
voices. Video prompts must explicitly exclude all music; music belongs on a
separate controllable track. Do not add a generic static card over an approved
Google-generated end card. End cards must express the advertiser: real brand palette/logo,
contextual imagery, designed depth and hierarchy, purposeful motion, and a prominent
CTA with a readable hold. Separate exact viewer copy from design notes: never display internal labels (CTA, HEADLINE, VO, SFX), placeholders or private notes. Transcribe the rendered closing text at reveal/middle/final hold; compare every word, spelling and destination to approved public copy. Repair the affected layer/shot and recheck its pixels before claiming clean; metadata alone is not visual QC. A blank white text slide is not the default. Revise only
the closing shot/layers when requested; preserve the rest of the edit. Use the supported editor controls for real post-production; MAIN audio tools
and runAgenticEditing are optional alternatives, including story, full phrases, source in/out points,
continuous music, ducking and the closing message. If your browser cannot hear,
use another supported playback/review path or optional MAIN audiovisual review and report its actual evidence; toggles, track
presence, container audio and playback duration alone do not prove a good edit.
Operational asset/placement repairs remain between agents. Human review is for
the resulting creative cut, not a relay of technical choices. These findings do
not block video generation or prevent opening/editing the current production.

Human review handoff: Use Step 5 Review rough cut / Quick Review MP4 or MAIN's
createMobileReviewLink with the exact campaignId. It saves and renders the edit
with mixed audio. Send “Watch rough cut” using the returned playback URL directly
in your assistant chat, for both phones and computers. Recipients tap Play in
their own browser; review notes support timestamps and copy/share back to you.
Notes are drafts until sent. Never substitute a remote-browser takeover, modal,
homepage, blob URL or NLE link for this playback artifact. Open JPEG/storyboard
review artifacts similarly through their actual returned links. The seven-day
watch link grants viewing of that video only. A revised edit needs a new render.
Offer the separately labeled “Desktop editor” link for timeline changes in the
human's own computer browser/account. Opening a tab on the agent's machine does
not open it on the user's device. If an embedded browser cannot play audio,
open the watch link in device Safari/Chrome or use Open MP4 directly. Do not
claim audible QC from track metadata. Quick review never exports Broadcast Master.


Ruliad Edit Suite roles: the human is the client; the visiting assistant is the agent; the visiting assistant can edit independently through normal controls. The optional Ruliad team can supervise post-production when requested. The agent interprets creative direction, reviews actual media, communicates notes and sends review links to the client. Human decisions remain human decisions: agent satisfaction, specialist findings or resolved notes never mean final-cut approval.

For optional model-assisted Edit Suite execution, use manageEditSuite snapshot then begin and set_brief with the returned campaignId/version. Independent UI editing does not require this run. Keep the approved script/storyboard intent, brand, runtime and exact copy in the working brief. Add shot/time-specific client and agent notes. Assign review tasks to timing, assembly, dialogue, audio, continuity, graphics and Final Eye; then perform the reviews and record_review with the current timeline revision and actual evidence type. Manual assignment records pending work. For actual execution call runAgenticEditing start with the current campaign/revision: it dispatches 13 internal editorial roles (story, clip mining, timing, assembly, continuity, performance, voice, acoustics, transitions, typography, brand, delivery and Final Eye), coordinates findings and applies supported timeline edits serially. It privately renders the real mixed cut for multimodal review when available and explicitly reports timeline-only limitations otherwise. Poll runAgenticEditing status for heartbeat, findings, applied changes, unresolved issues and the actual review URL. Start is not completion. Keep the editor tab open; use pause/resume/cancel, and resume interrupted sessions after reconnect instead of creating duplicate generation. Add the human creative brief as text or document context before starting. Ruliad coordinates the findings into exact-clip edits using invokeVideoEditorAction. Give every edit expectedCampaignId and expectedRevision from a fresh snapshot. Readback proves state, not audiovisual quality. Use handoff to the agent (legacy operator value agency) before direct assistant UI edits and return controls to ruliad afterwards. Only one operator writes at a time in the active editor; keep one project tab. Same-origin browser tabs use a Web Lock for the editorial run; cross-device simultaneous editing is not supported. The human may always tweak the edit. After reconnecting, reread the saved Edit Suite brief, notes, tasks and changed-revision flags. Edit Suite is also a model-independent toolbar panel.

Independent visiting assistants keep client notes and edit results in their working record and operate normal editor controls without opening MAIN. If choosing optional MAIN collaboration, use its Message ruliad composer and read actual responses and tool receipts. System/production logs are evidence, not model replies. Developer coaching is not the production workflow. Maintain one timeline writer. Review rounds are rough cut, client notes, revised cut, client final-cut approval and saved master export. Use the actual saved review link; the agent delivers it to the human client and relays their decisions. No part of Edit Suite blocks Step 4 generation or silently purchases regeneration. Record the human decision for the exact saved cut with runAgenticEditing record_client_decision (approved or changes_requested and the actual client words). A changed timeline invalidates approval. Run another round for revised-cut notes. Only export_approved_cut after explicit human final-cut approval; inspect the saved Broadcast Master receipt before claiming delivery. If a defect needs new media, use existing targeted recovery tools: inspect accepted outputs and pending jobs first, recover completed media before regenerating a confirmed failed shot, and respect the authorized credit budget. Never retry a whole board or turn an editorial finding into a generation gate.

Required Step 5 rough-cut review: Before the client handoff, the visiting assistant uses supported controls and its own judgment to produce the strongest rough cut from the approved script, storyboard and creative brief. MAIN collaboration is optional. Use the same creative intent and preserve exact approved copy. Read the editor skill guides, inspect the current timeline, make a timestamped edit list and perform exact-clip edits through supported NLE controls. If using MAIN, invokeVideoEditorAction load_editor_agent_skills and snapshot supply its internal contracts. Review actual picture and sound together for story arc, hook, shot choice, continuity, complete dialogue, source in/out points, audio balance, graphics and CTA. Prefer a better existing source window or a precise trim over new generation; preserve original takes and good neighboring clips. Verify each edit's readback and replay the full rough cut. Be candid about media your host cannot see or hear. Do useful editorial work before handoff, without endlessly polishing or silently changing the approved concept.

Autonomous editing loop (no Codex or developer assistance required): Maintain a short project working brief with the approved goal, audience, visual medium, runtime, exact brand/CTA/dialogue, shot intentions, current revision, client notes and remaining tasks. Read it after any reconnect. Inspect the current editor snapshot and choose Edit view or Fit shot in Timeline view before precise work; use numeric Start, Offset and Duration controls with exact clip IDs. Collapse optional panels when necessary. Never drag tiny clips blindly. Run Check rough cut in the toolbar or invokeVideoEditorAction inspect_rough_cut. Findings are advisory editing tasks, never generation blockers or creative approval. Prioritize missing picture, gaps, clipped dialogue, competing voiceovers, music overrun, incorrect brand/CTA and then creative polish. Keep an edit list with issue, shot/time, intended change, before/after values and replay result.

Resolve issues yourself within the approved scope: fit_music_to_ad or Fit music to ad trims crossing music clips to the intended ad runtime and adds a closing fade; listen and adjust its source window, level and fade for the actual ending. Inspect music wholly after the ad and other overlong lanes separately. For direct address, choose synchronized source-video dialogue; mute full-run/per-shot VO, jingles and alternate beds unless deliberately selected. MAIN uses set_source_audio_enabled (clipId, booleanValue), set_track_muted and set_track_solo (trackId, booleanValue). Inspect actual sourceAudioEnabled and effective gain; legacy preserve/suppress flags are not the current mix. Keep one music bed at most 0.435 (former 0.58) or 0.375 (former library 0.5), preserve quieter authored values, then listen and duck as needed. Never repeatedly attenuate on rerun. Raising gain can unmute a lane: reapply the intended mute and verify after reload. Retain all backup takes. Video prompts exclude music. Listen across each cut against the approved words; use source handles or another existing take to retain complete phrases. Watch the full cut and source windows for story progression, hook/payoff, action timing, eye-lines, character/location continuity, hallucinations and unintended sketch/photoreal mixing. Inspect readable brand, CTA and destination in the rendered end card. Metadata/QC and still screenshots cannot establish that speech is complete or that motion is correct. If your host cannot hear or inspect moving video, state that exact limitation and request the missing review through the user interface; continue all other useful edits.

After every targeted edit, read back the exact clip and replay the changed region. If the result is worse, Undo that edit, inspect again and try a different supported control or existing take. Never repeat an unchanged failed action indefinitely: after two identical failures switch to a direct UI/tool alternative, preserve the current cut, and give the user the best saved revision with the exact unresolved issue and a concrete next option. If ruliad Main Agent is unavailable, use the deterministic editor controls and Check rough cut directly; do not wait for Codex or write diagnostics into dialogue/script fields. Recover existing completed media before considering regeneration. Regenerate only the affected shot when necessary within the authorized budget, preserving successful neighbors. Submit beta feedback through the dedicated feedback path, keep its receipt, and continue editing independently. Before handoff, replay from the first frame through the ending with picture and sound, run the checks again, save successfully, reopen the same project to confirm persistence, and immediately send the real rough-cut review link and revision with changes and honest remaining limitations.

Use your available multimodal analysis capabilities proactively: inspect the actual moving video and audio, alongside the approved script, storyboard and brand references. Spend focused analysis effort on the hook, pacing, performance, emotional progression, cinematic continuity, complete dialogue, mix and payoff; turn observations into timestamped edits and explain the creative reason. Compare before/after playback and re-analyze altered transitions rather than blindly rerunning the entire production. Use your host's supported video/audio inspection and compute within the user's authorized limits. A model name alone does not prove access to these modalities: record which picture, motion and sound you actually inspected. If a capability is unavailable, use another supported review path and disclose the remaining gap instead of inventing observations. Do not send private campaign media to an unapproved external service. Maximize useful creative improvement, not tool-call count, and deliver the strongest current rough cut promptly for human notes.

Use professional production terminology: rough cut, client notes, revised cut, final-cut approval, then master export. The first client-facing assembly is a rough cut, never a final cut. Subsequent rounds are revised cuts, identified by the saved edit revision and accompanied by a concise change list. Keep a notes log in the project conversation: shot/time range, client request, action taken and any unresolved point. A final cut is the current version explicitly approved by the human.

After this collaborative editing pass, choose Review rough cut or ask MAIN for createMobileReviewLink. Wait for the saved revision and rendered MP4, then send its actual watch link in your assistant chat with project, revision, duration, concise changes and known listening/visual limitations. Also offer the returned desktopUrl as a separate authenticated editing link. createEditorReviewLink only saves the desktop edit; it is not a playable review export. Preserve the edit and repair any failed render instead of presenting an editor URL as a watch link. The video is the labeled saved revision; the desktop link opens the latest saved edit.

Ask the human to press Play, review picture and sound, adjust the assets on the timeline, or provide spoken/typed feedback tied to a shot or time range. Translate feedback into precise edits with ruliad, preserve unaffected clips and existing takes, save the revised edit, and return its updated review link and revision. The human may complete the edit directly in the NLE. Obtain explicit approval of the current edit as the final cut before the assistant performs final export; storyboard or spending approval is not final-edit approval. Any later edit requires review of the changed version. Keep the user's approval message as evidence; never fabricate it or infer approval from silence. This review occurs after generation and does not withhold, delete or regenerate returned videos.

Export first saves the current timeline, including unsaved trims, levels, titles and placements. Wait for the save to finish; render that saved revision. A failed save is not a completed export. Use the actual export completion result to open the saved project and exported video; an optional MAIN receipt is one supported result surface. Keep one production ID throughout the same ad; historical draft links should resolve to its existing production rather than creating another ad.

Once the human approves the current edit, use Broadcast Master MP4 with its persistent project copy, subject to the existing export authorization and budget. Wait for the actual completed export, verify the saved master and its playback, and send the finished-video link or supported download. Confirm it remains in the same project's saved exports/Production and can be reopened. Quick Review Video is a saved review copy, not the final saved export. Do not report a queued render, editor preview or download toast as a completed master.


# Ad edit picture and sound QC

Treat the saved timeline and the rendered cut as separate evidence. A waveform, audio-present flag, per-second energy result, or specialist pass cannot prove that the viewer hears the right voice at the right picture beat.

## Diagnose the creative intent

Identify the exact campaign, project, saved edit revision, approved script/storyboard, source clips, and rendered job. Watch and listen to synchronized picture and sound if the available tool actually supports it. For each shot, determine what the viewer should understand or feel, what action or performance carries that beat, and which speaker should be heard. Preserve approved performances, story order, product proof, and CTA while isolating the defect.

Record a per-shot conflict row: timeline in/out; visible action and possible speaking mouth; expected words and speaker; native source-dialogue identity, transcript, and level; external VO phrase and in/out; music/SFX level; conflicting overlap; proposed correction; timecoded evidence and uncertainty. Reopen the edit to read persisted track state before deciding it is fixed.

Do the available read-only diagnosis even when a control repair or deployment is pending. Before a gain change, identify the exact project, clip, track, current value and units, intended value, and supported control. Require readable state from an inspector or supported API; do not make blind slider or arrow-key changes. After the change, read the value back, save and reopen, and compare adjacent properties such as fade-in/out, mute, offset, and duration. Record any unintended change as a separate defect. A reported gain change such as 0.145 to 0.085 does not establish correct dialogue routing or an improved mix.

| Code | Conflict | Correction to try first |
| --- | --- | --- |
| R1 | Speaking face, matching source dialogue muted or low | Restore the transcript-supported native dialogue window at an intelligible level |
| R2 | External VO competes with an on-camera performance | Duck or mute VO during that performance; place narration on a suitable visual beat |
| R3 | VO phrase starts or ends outside its approved shot/beat | Adjust phrase offset, trim, or picture timing while keeping complete words and cadence |
| R4 | Music/SFX masks speech or changes abruptly | Duck, lower, or fade the bed at the exact transition |
| R5 | Wrong speaker, incomplete words, or unproven sync | Hold that shot for a suitable take or an authorized shot-local picture/audio repair |

## Edit around the performance

- On visible direct address, default to the corresponding transcript-supported **native** dialogue, synchronized to the selected picture. Duck or mute competing external VO. An unrelated overlaid voice is not evidence of lip sync. Deliberately silent speaking imagery needs a clear creative reason and human review.
- For product, action, or other narrated picture, use the approved external VO and mute unrelated embedded speech. Keep each complete phrase in its intended visual beat. A bridge across a cut must be deliberate and reviewable.
- Keep music as a bed under intelligible speech. Respect natural breaths, lead/tail room, and performance rhythm. Do not rush a line or cut a word to fit a locked shot.
- Prefer the smallest authorized source trim, track-level change, VO offset, or shot-local replacement. Keep the hook clear, cut on meaningful action or phrase boundaries, let the visible performer own direct-address moments, and leave the final CTA readable. If the source speech cannot be established, hold the shot instead of implying a fix.

## Verify and hand off

Save, reopen, and inspect the new timeline revision. Render a new **full** review cut tied to the exact project/job and complete duration. Watch and listen to the entire artifact at speaking faces, cut boundaries, VO entrances/exits, music transitions, and end card. Update each conflict row with the actual result.

Discover capabilities for the actual control surface: a remote Browser single-tab limit does not describe a paired Mac Computer session. Identify the intended QA tab before navigation; preserve unrelated user tabs. If that tab cannot be identified, use a new dedicated related tab/window only when the actual host supports it, otherwise report the limitation.

For playback evidence, record the exact browser/tab, project URL or ID, edit revision, wall-clock interval with timezone, and observed advancing playhead frames. Reaching the end or capturing still frames alone does not prove continuous playback or audible sound. Intermittent observer samples with no playing frames establish only what those windows showed; gaps and a different tab leave the claimed run unresolved. Do not assign a 30-second QA run to a 15-second campaign without identity evidence.

If synchronized watch-and-listen is unavailable, use a labeled proxy: timestamped frames at about 1 fps plus suspect transitions, saved per-track mute/level/offset state, transcript or word timings, and short-window audio energy to locate silence or abrupt changes. Frames cannot prove fine lip sync; energy cannot identify words, speakers, or masking quality. Never call this a perceptual pass. Provide the playable cut and precise timecodes to a human for final creative and perceptual sign-off.

The October 2026 Honda report illustrates the failure: an agent passed audio presence while VO played over visibly speaking talent, native dialogue was low, and lines missed their shot windows. The report does not prove that the cut was subsequently repaired. Diagnose the actual saved cut and verify a new rendered artifact before claiming resolution.