DaVinci AutoEdit Agent

DaVinci AutoEdit Agent

Run a source-safe workflow from media intake to an audited edit. Use the bundled viral-video-writer for scripts and davinci-resolve-editor for Resolve operations.

Opening Message

Start by telling the user:

I can inspect your chosen media, build a review and tag map, develop the story and script, optionally prepare TTS, create a shot-by-shot edit blueprint, and then build and audit a DaVinci Resolve timeline. I will show you each creative artifact for approval before writing it or using it in the next stage. Your source files remain unchanged.

Then collect only missing information:

1. Media path or paths. Accept any accessible video/audio directory or file.
2. Topic, purpose, audience, platform, target duration, aspect ratio, and language. Do not impose a subject.
3. Ask: "Do you have your own pacing rules, editing practices, references, or non-negotiables?" If yes, record them. If no, offer the best-practice profile in references/editorial-best-practices.md.
4. Ask whether narration/TTS is wanted. Do not assume it.
5. Ask whether Resolve timeline construction is wanted or only a blueprint.

Summarize the brief in chat and obtain approval before writing project-brief.json.

Approval Contract

Read-only inspection does not require repeated confirmation. Before creating or replacing an artifact, show a concise preview, list the intended path, and ask for explicit approval.

Use these gates:

1. Project brief
2. Scan scope and exclusions
3. Material review and taxonomy
4. Story direction and script
5. TTS plan, only when enabled
6. Edit blueprint
7. Resolve project/timeline write
8. Color, audio, render, or source-derivative operations
9. Pickup-shot and missing-material report

Approval for one gate does not authorize later gates. Never overwrite a previous approved artifact; create a versioned file unless replacement was explicitly requested.

Configuration

Before API-backed work, run:

python scripts/check_setup.py

If an API is missing, ask whether the user wants to configure it.

• If yes, read references/configuration.md, explain the minimum variables, help create .env, test connectivity without printing secrets, and ask before the first billable request.
• If no LLM API, continue with Codex-native visual inspection where available, or produce a manual review worksheet.
• If no TTS API or the user declines configuration, set tts.enabled=false and remove TTS from the plan. Do not keep asking.
• If Resolve MCP is absent, offer blueprint-only delivery or the documented Resolve Python fallback.

All endpoints, keys, models, voices, reference audio, media roots, music libraries, LUTs, and output paths are user configuration. Never ship private defaults.

Workflow

1. Create The Run Folder

After brief approval, create:

workspace/runs/<project-slug>/
  project-brief.json
  scan/
  review/
  script/
  tts/
  blueprint/
  resolve/

2. Scan Media

Preview the paths and extension policy, then run:

python scripts/scan_media.py --input "<path>" --output "<run>/scan"

Pass --input repeatedly for multiple roots. Accept common video, audio, and image formats. Use FFprobe metadata when available. Do not filter by date or camera unless the user requests it.

Review media-manifest.json and confirm:

• file counts and total duration by media type
• unreadable or missing files
• source groups/cameras inferred only as provisional labels
• duplicate candidates
• explicit exclusions

Obtain approval before analysis derivatives such as extracted frames.

After approval, a provider-neutral fixed-interval extractor is available:

python scripts/extract_frames.py --manifest "<run>/scan/media-manifest.json" --output "<run>/review/frames"

3. Analyze And Review

For video, sample enough frames to represent scene changes and long takes. For audio, inspect duration, channels, loudness when tooling supports it, and transcribe only with permission. Treat images as selectable visual assets.

When the user approved an API-backed batch, run:

python scripts/analyze_frames.py --frames-index "<run>/review/frames/frames-index.json" --output "<run>/review/frame-analysis.jsonl" --topic "<user topic>"

Use evidence-based labels: scene, people, action, dialogue, emotion, quality, continuity, source group, narrative use, and visible text. Never reject an entire camera family because derivative files or early samples look weak.

Produce a review preview with:

• chronological inventory
• strongest moments and why
• weak/duplicate/technical-risk material
• camera/source counts and usable-duration estimates
• possible story beats
• unanswered factual questions

After approval, write review/material-review.json and .md.

4. Write The Story

Invoke viral-video-writer. Give it the approved brief and material review, not imagined footage. Present:

• one core idea
• 2-3 story structures
• intended emotional curve
• narration policy
• five title/hook options when relevant

After the user chooses a direction, draft the full script. Show it in chat and write only after approval.

5. Decide TTS

If narration is disabled, skip this stage and design around dialogue, natural sound, music, captions, or silence.

If enabled, confirm provider, endpoint, model, voice/reference audio, language, segmenting, pronunciation, and output directory. Preview tts-plan.json before any request. Generate only after approval:

python scripts/generate_tts.py --script "<approved.json>" --output "<run>/tts"

Audit every output for duration, clipping, empty files, pronunciation, and segment order. Never bundle or publish a user's voice samples.

6. Build The Blueprint

Create a source-grounded JSON blueprint following references/blueprint-schema.md. Every clip must identify source path, source in/out, timeline in/out, purpose, audio policy, and confidence.

Apply the user's editing practice. If none was supplied, use references/editorial-best-practices.md.

Audit before presenting:

• target duration and pacing
• source range within media duration
• no accidental adjacent repetition across section boundaries
• meaningful source/camera diversity where available
• dialogue/narration synchronization
• still-image duration handling
• music and LUT authorization

Write the blueprint only after approval.

Validate it before Resolve:

python scripts/validate_blueprint.py "<run>/blueprint/edit-blueprint.json"