Automate This

# Automate This

Analyze a screen recording of a manual process and build working automation for it.

The user records themselves doing something repetitive or tedious, hands you the video file, and you figure out what they're doing, why, and how to script it away.

## Prerequisites Check

Before analyzing any recording, verify the required tools are available. Run these checks silently and only surface problems:

```bash
command -v ffmpeg >/dev/null 2>&1 && ffmpeg -version 2>/dev/null | head -1 || echo "NO_FFMPEG"
command -v whisper >/dev/null 2>&1 || command -v whisper-cpp >/dev/null 2>&1 || echo "NO_WHISPER"
```

- **ffmpeg is required.** If missing, tell the user: `brew install ffmpeg` (macOS) or the equivalent for their OS.
- **Whisper is optional.** Only needed if the recording has narration. If missing AND the recording has an audio track, suggest: `pip install openai-whisper` or `brew install whisper-cpp`. If the user declines, proceed with visual analysis only.

## Phase 1: Extract Content from the Recording

Given a video file path (typically on `~/Desktop/`), extract both visual frames and audio:

### Frame Extraction

Extract frames at one frame every 2 seconds. This balances coverage with context window limits.

```bash
WORK_DIR=$(mktemp -d "${TMPDIR:-/tmp}/automate-this-XXXXXX")
chmod 700 "$WORK_DIR"
mkdir -p "$WORK_DIR/frames"
ffmpeg -y -i "<VIDEO_PATH>" -vf "fps=0.5" -q:v 2 -loglevel warning "$WORK_DIR/frames/frame_%04d.jpg"
ls "$WORK_DIR/frames/" | wc -l
```

Use `$WORK_DIR` for all subsequent temp file paths in the session. The per-run directory with mode 0700 ensures extracted frames are only readable by the current user.

If the recording is longer than 5 minutes (more than 150 frames), increase the interval to one frame every 4 seconds to stay within context limits. Tell the user you're sampling less frequently for longer recordings.

### Audio Extraction and Transcription

Check if the video has an audio track:

```bash
ffprobe -i "<VIDEO_PATH>" -show_streams -select_streams a -loglevel error | head -5
```

If audio exists:

```bash
ffmpeg -y -i "<VIDEO_PATH>" -ac 1 -ar 16000 -loglevel warning "$WORK_DIR/audio.wav"

# Use whichever whisper binary is available
if command -v whisper >/dev/null 2>&1; then
  whisper "$WORK_DIR/audio.wav" --model small --language en --output_format txt --output_dir "$WORK_DIR/"
  cat "$WORK_DIR/audio.txt"
elif command -v whisper-cpp >/dev/null 2>&1; then
  whisper-cpp -m "$(brew --prefix 2>/dev/null)/share/whisper-cpp/models/ggml-small.bin" -l en -f "$WORK_DIR/audio.wav" -otxt -of "$WORK_DIR/audio"
  cat "$WORK_DIR/audio.txt"
else
  echo "NO_WHISPER"
fi
```

If neither whisper binary is available and the recording has audio, inform the user they're missing narration context and ask if they want to install Whisper (`pip install openai-whisper` or `brew install whisper-cpp`) or proceed with visual-only analysis.

## Phase 2: Reconstruct the Process

Analyze the extracted frames (and transcript, if available) to build a structured understanding of what the user did. Work through the frames sequentially and identify:

1. **Applications used** — Which apps appear in the recording? (browser, terminal, Finder, mail client, spreadsheet, IDE, etc.)
2. **Sequence of actions** — What did the user do, in order? Click-by-click, step-by-step.
3. **Data flow** — What information moved between steps? (copied text, downloaded files, form inputs, etc.)
4. **Decision points** — Were there moments where the user paused, checked something, or made a choice?
5. **Repetition patterns** — Did the user do the same thing multiple times with different inputs?
6. **Pain