Clip Feedback — Prompt Intent vs Reality

What this does

Takes a video clip plus the prompt that generated it, analyzes what the clip actually looks like (via Gemini 2.5 Pro), and returns:

• Rule violations — character names in prompt, 4th wall breaks, missing focal length, safety trigger stacking
• Intent gaps — emotion mismatch, world palette drift, shot type wrong, energy level off
• Quality issues — text overlays, CGI artifacts, motion blur, missing face
• Specific prompt fixes — exact corrections to improve the next generation

Usage

Quick rule check (no API call — instant)

python3 scripts/prompt_feedback.py \
  --clip /path/to/clip.mp4 \
  --prompt "the generation prompt used" \
  --rules-only

Full analysis (calls Gemini 2.5 Pro — ~10s)

python3 scripts/prompt_feedback.py \
  --clip /path/to/clip.mp4 \
  --prompt "the generation prompt used" \
  --source veo

Supported sources: veo, seedance, kling, runway, or any label.

From a prompt file

python3 scripts/prompt_feedback.py \
  --clip /path/to/clip.mp4 \
  --prompt-file /path/to/prompt.txt \
  --source seedance

JSON output (for piping to other tools)

python3 scripts/prompt_feedback.py \
  --clip /path/to/clip.mp4 \
  --prompt "..." \
  --json

When the user says "check the clip"

1. Ask which clip (path) if not obvious from context
2. Ask for or find the prompt that generated it
3. Run --rules-only first for instant feedback
4. If the user wants deeper analysis, run the full version
5. Present the feedback summary and suggested fixes
6. Offer to apply the fixes and regenerate

What it checks (canonical rules)

Configuration

Configure your canonical rules in reference/prompt-rules.md. Each rule gets:

• A pattern (regex or keyword list) to detect violations in prompts
• A fix template (the correction to suggest)
• A severity (block / warn / info)

Dependencies

• Python 3.10+
• google-generativeai for Gemini analysis
• ffmpeg on PATH for frame extraction

Setup

pip install google-generativeai pillow
export GEMINI_API_KEY=your_key_here