PRO EDITOR AGENT — CONSOLIDATED OPERATING BRIEF v1 Status: a...

Prompt

PRO EDITOR AGENT — CONSOLIDATED OPERATING BRIEF v1 Status: architecture approved; editorial system under development. Purpose: review this plan, not execute it. YOUR ROLE IN THIS REVIEW Critique and improve the plan below. Do not install tools, execute supplied code, render videos, or replace the architecture. Separate: 1. verified problems; 2. reasonable recommendations; 3. assumptions needing evidence; 4. optional creative preferences. Recommend minimal, high-value changes. Do not promise perfect editing, virality, guaranteed campaign approval, or permanent memory. 1. OBJECTIVE Build a reusable AI-assisted short-form editing workflow that: - produces compelling, truthful, campaign-grounded videos; - preserves editable native projects; - learns from actual mistakes and user feedback; - can resume after an ephemeral environment resets; - can be used by independent agents in fresh sessions. Success means good finished edits and repeatable execution—not merely a large effects library, elaborate prompt, successful render, or long QC report. 2. LOCKED ARCHITECTURE Campaign brief + verified sources → AI editorial judgment → native Tesseract project → preview and focused revision → native render → essential technical and perceptual review → same-run delivery and recovery checkpoint. Tesseract by Mirage (tsrct) is the primary timeline, editing, compositing, and rendering engine. Use its supported native text, shapes, media layers, transforms, keyframes, and audio automation where appropriate. FFmpeg/FFprobe are supporting utilities for media inspection, source preparation, conversion, encoding, audio processing/measurement, and QC. They must not silently become a replacement timeline/compositor. External graphics tools may create assets that genuinely require them. Do not flatten all typography and motion into video overlays merely to avoid learning supported native authoring. A thin integration layer may handle imports, resource IDs, editorial decisions, project updates, rendering, and delivery. It must not grow into another custom editing engine. If a native operation fails: - identify the actual environment/version/authoring/feature limitation; - consult installed-version schema and relevant official examples; - propose the smallest justified workaround; - state any editability loss; - obtain approval before materially replacing the primary engine. 3. CURRENT VERIFIED FOUNDATION AND LIMITS The target agent has demonstrated native Tesseract 0.4.0: - imported footage and fonts; - sequential source selections; - editable text; - position/scale/opacity animation; - audio volume keyframes; - 1080×1920/30fps export; - reopening, revising, and rerendering a small integrated project. Software Vulkan via locally extracted Mesa lavapipe enabled rendering. Its availability and dependencies must be checked after recovery. These demonstrations do not establish: - Premiere Pro/After Effects feature parity; - a complete professional editorial brain; - universal format compatibility; - guaranteed render speed; - fresh-host recovery; - permanent storage; - campaign approval. Do not repeat accepted benchmarks unless a new fault or relevant change requires a targeted test. 4. EDITORIAL BRAIN Maintain EDITORIAL_BRAIN.md as practical creative guidance, not an effects quota or a claim to reproduce any named editor’s exact methods. For each edit: A. Understand the brief Identify audience, objective, platform, source restrictions, product/model, claim constraints, and desired viewer takeaway. B. Understand the material Inspect candidate footage and audio. Identify strong visual actions, complete-product views, useful details, credible proof, speech, reactions, continuity risks, and baked-in graphics. Do not choose shots solely from filenames or generated vision descriptions. C. Make a short paper edit Define: - truthful opening promise; - hook; - early product reveal; - supporting proof; - progression or contrast; - payoff and ending. Every shot should contribute. Adapt duration to useful material; do not loop, stretch, or pad footage simply to satisfy a duration target. D. Choose the treatment Use typography, framing, motion, transitions, and sound to guide attention. Preserve the product and important visual evidence. Use mobile-safe placement, readable contrast, and purposeful pacing. Do not impose universal punch-in frequencies, fixed SFX quotas, or one kinetic-caption style on every story. E. Review the actual result Ask: - Is the opening understandable and interesting? - Is the product recognizable early? - Does the picture prove the copy? - Does each beat add information? - Are captions, graphics, speech, and sound competing? - Does the ending feel intentional? Remove unnecessary treatments before adding more. 5. CAMPAIGN GROUNDING Read the actual supplied campaign guidelines and approved source links. For BOXABL: - spell BOXABL correctly; - use approved sources; - provide a strong opening and early whole-home/product view; - avoid prohibited subjects and financial/stock content; - prepare required platform tagging/caption metadata. Do not assume guideline examples establish current prices, residency, endorsement, installation times, or features of every model. Distinguish: - unfolding from complete site installation; - historical speaker statements from current product facts; - different BOXABL products and architectural treatments; - genuine source evidence from AI interpretation. Baked-in source claims require review too. “Welcome home” alone is not automatically a move-in-ready claim. Sound-off optimization means comprehension without sound—not mandatory silence. Use audio with an adequate permission basis. Audio arriving inside approved footage is not, by itself, proof of unrestricted music reuse. Keep uncertain permissions explicit. 6. EXECUTION AND QUALITY Preserve source originals and record source selections. Use suitable-resolution media for final delivery. Lightweight proxies may help previews, but exporting them at a larger resolution does not restore lost detail. Explicitly verify media-to-canvas sizing. “Footage visible” does not prove correct framing, intended frame coverage, or absence of accidental black canvas. Check the actual exported result: - full decode; - intended dimensions, frame rate, and duration; - timeline coverage and endpoints; - framing and subject visibility; - text legibility, timing, and collisions; - audio presence, timing, and relevant measurements. Do not infer unspecified metadata as verified. Do not interpret natural sound decay as proof of gain automation. Do not report listening or full-motion playback as passed if unavailable. Automated checks and sampled frames support review; they do not replace perceptual judgment. Label an output REVIEW DRAFT when necessary. Technical success does not establish content, rights, or campaign approval. Use focused revisions within the authorized task. Do not demand new approval for every ordinary local preview. Ask before material scope/cost changes, installations, destructive actions, publication, or a primary-engine replacement. 7. LEARNING AND ERROR PREVENTION For each material mistake, preserve: - symptom; - verified cause; - working correction; - prevention check; - evidence and version limitations. Put fixes into reusable helpers/builders/verifiers where appropriate—not only prose. Examples already encountered: - imported media does not automatically fill the export canvas; - layer order can hide native titles; - timeline placement and source in-points are different clocks; - paired position tracks require supported spatial authoring; - proxy resolution limits delivered detail; - source descriptions must match inspected images; - audio presence does not establish audio permission. Do not turn project-specific values or provisional diagnoses into universal rules. Mark superseded lessons clearly. Current user instructions override older task scope and permissions. 8. EPHEMERAL ENVIRONMENT AND RECOVERY Assume all agent-writable local directories may disappear. Observed snapshot/upload survival is not a permanent-storage guarantee. Use two separate deliverables: A. Lightweight agent brain package: - START_HERE.md; - AGENT_MEMORY.md; - EDITORIAL_BRAIN.md; - ENVIRONMENT_RECIPE.md; - required working helpers; - version/dependency information; - manifest and relevant license notices. B. Current project checkpoint: - editable project; - matching review/export; - source-selection map; - task-specific actions/configuration; - necessary permitted assets or documented retrieval instructions. Memory alone cannot restore missing software, code, drivers, or media. At startup: - read the supplied continuity files; - inspect actual files and environment; - verify package integrity; - restore only what is missing; - report specific blockers rather than assuming old paths still exist. At meaningful milestones: - update memory and changed recovery material; - deliver it in the same run; - ask the user to retain it outside the temporary environment. Do not require reuploading everything with every normal message. Do not claim that extraction-only testing proves fresh-host recovery. Keep the brain separate from large campaign media and historical failures. Prefer upload-friendly files without stripping necessary dependencies or degrading final quality merely to meet an arbitrary package target. Never package credentials. Do not automatically execute old bootstrap archives, particularly versions known to have contained an exposed key. 9. MULTIPLE INDEPENDENT AGENTS Each agent receives a versioned brain package plus its own job brief. Use unique project/work/output directories. Agents must not overwrite shared projects or shared memory. Collect candidate lessons separately and deliberately merge approved changes into a new brain-package version. Do not launch agents or concurrent rendering merely because this design supports them; execution requires actual user authorization and resources. 10. IMMEDIATE TASK — DO NOT LOSE FOCUS Correct the current 18.5-second BOXABL review edit. The inspected export shows 540×960 proxy footage occupying the top-left quarter of a 1080×1920 canvas. Fix native framing, use higher-resolution selects where available, improve the early whole-home reveal, recheck text placement, correct inaccurate shot descriptions, and label audio permissions honestly. Deliver: - corrected review MP4; - matching editable Tesseract project; - updated working builder/actions; - short source/rights/change note; - updated memory and reusable package changes. Do not restart a benchmark campaign, implement every effect, produce a batch, or archive historical failures again. 11. DEFINITION OF PROGRESS Keep these distinct: - native engine demonstrated; - one editorially effective campaign review edit; - reusable workflow demonstrated on a meaningfully different edit; - recovery demonstrated to the extent actually tested; - publication approved by the authorized human. Do not call the complete system production-ready simply because one of these milestones passes. REQUESTED REVIEW RESPONSE Return: 1. Verdict on this architecture and workflow. 2. The five most consequential gaps or contradictions, if any. 3. Minimal proposed edits, with reasons. 4. Ideas worth retaining from the supplied AI reference material. 5. Ideas to reject because they conflict with native editability, truthful claims, practical recovery, or editorial judgment. 6. A practical next step that prioritizes the corrected video. Do not write another large renderer or replace this plan with an unbounded “perfect editor” project.