m2ai-portfolio Video-to-Skill — Screen Recording to Automated Skill Pipeline
Turn a screen recording of any manual process into a fully structured Claude Code skill. This captures tacit knowledge (cursor movements, under-breath commentary, style preferences, iteration patterns) that written prompts miss.
Prerequisites
gemini-api-devskill installed globally- Gemini API key in
~/.env.sharedasGOOGLE_API_KEY skill-creatorskill available (for final skill generation)- Screen recording software (Loom, OBS, native OS recorder)
ffmpegor Python video compression available (for large recordings)
Phase 1: Video Preparation
- User provides a screen recording file (MP4, WebM, MOV).
- Check file size. If over 50MB, compress before upload:
python3 -c "
import subprocess, sys, os
inp = sys.argv[1]
out = os.path.splitext(inp)[0] + '_compressed.mp4'
subprocess.run([
'ffmpeg', '-i', inp, '-vcodec', 'libx264', '-crf', '28',
'-preset', 'fast', '-vf', 'scale=-2:720', '-an', out
], check=True)
print(f'Compressed: {os.path.getsize(out) / 1e6:.1f}MB')
" /path/to/video.mp4
- Confirm the compressed file is under 50MB. If still too large, increase CRF or reduce resolution.
Phase 2: SOP Extraction via Gemini
Send the video to the Gemini video understanding API using the gemini-api-dev skill.
Prompt to use for extraction:
Take this screen recording and produce the most comprehensive, detailed SOP of the entire process you see me walk through. Account for:
- Every step taken, including navigation and UI interactions
- All spoken commentary, especially offhand remarks and under-breath observations
- Style preferences expressed (what the user likes/dislikes and why)
- Iteration patterns (what was changed, how many times, what triggered changes)
- External references visited (websites, examples, inspiration sources)
- Decision points where the user chose one option over another and why
- Any tacit knowledge that wouldn't normally be written down
Structure the SOP as numbered phases with sub-steps. Flag "tacit knowledge nuggets" separately at the end.
Save the extracted SOP to a working directory (e.g., /tmp/video-to-skill/sop-{name}.md).
Phase 3: SOP Review (Human-in-the-Loop)
Present the extracted SOP summary to the user. Ask:
- Does this accurately capture your process?
- Any steps missing or misinterpreted?
- Any tacit knowledge nuggets that are wrong or need emphasis?
Incorporate feedback before proceeding.
Phase 4: Skill Generation
Invoke the skill-creator skill with the reviewed SOP as input.
Key instructions for skill-creator:
- Use the SOP as the ground truth for the skill's phases
- Preserve all tacit knowledge as explicit rules or preferences in the skill
- Add
AskUserQuestioncheckpoints at decision points identified in the SOP (minimal HIL) - Include source attribution in the skill frontmatter
- Style preferences from the SOP become hard rules in the skill, not suggestions
Phase 5: Validation
- Read the generated SKILL.md
- Verify all SOP phases are represented
- Verify tacit knowledge nuggets are encoded as rules
- Use
claude-code-guideagent to validate skill structure - Test invocation in a fresh session
Cost Management
| Recording Length | Estimated Gemini Cost | Tip |
|---|---|---|
| Under 5 min | < $0.05 | Send as-is |
| 5-15 min | $0.05-0.20 | Compress to 720p |
| 15-40 min | $0.20-0.80 | Compress aggressively (CRF 30+, 480p) |
| Over 40 min | $0.80+ | Split into segments or compress hard |
When NOT to Use This
- Process is purely text-based with no visual/UI component (just write the skill directly)
- Recording is mostly idle time or waiting (trim first)
- The task is a one-off that won't be repeated
Source Attribution
Technique from Mark Kashef: "You've Never Made a Claude Code Skill Like This" (2026-03-23) https://www.youtube.com/watch?v=hTWxGSsGDZU