Google Image Search Skill

---
name: google-image-search-2d611b6356b1
description: Search and download images via Google Custom Search API with LLM-powered selection. This skill should be used when finding images for articles, presentations, research documents, or enriching Obsidian notes with relevant visuals. Supports simple queries, batch processing from JSON config, automatic config generation from terms, and full note enrichment with automatic image insertion below headings.
---

# Google Image Search Skill

Search for images using Google Custom Search API with intelligent scoring and LLM-based selection.

## When to Use

- Finding images to illustrate technical articles or research
- Adding visuals to presentations
- Enriching Obsidian notes with relevant images
- Batch image search for multiple topics
- Generating image search configs from plain text lists

## Requirements

- Google Custom Search API key and Search Engine ID
- OpenRouter API key (for LLM selection)
- llm CLI installed at `/opt/homebrew/bin/llm`

Store credentials in `.env`:
```
Google-Custom-Search-JSON-API-KEY=your_key
Google-Custom-Search-CX=your_cx
OPENROUTER-API-KEY=your_openrouter_key
```
(The `OPENROUTER_API_KEY=...` spelling works too — see the plugin note below.)

### OpenRouter / llm-openrouter caveat

The `llm-openrouter` plugin only registers its models when it can read a key named
`OPENROUTER_KEY` (env var or `llm keys set openrouter`). With only
`OPENROUTER_API_KEY` set, `llm models list` shows **zero** OpenRouter models and every
`openrouter/...` name fails with `Unknown model` — the script then silently falls back to
keyword scoring. `llm_select.py` sets both names for the subprocess, so this is handled;
if selection ever regresses, check `OPENROUTER_KEY` first:

```bash
llm models list | grep -c openrouter   # 458+ means the plugin registered its models
```

## Modes of Operation

### 1. Simple Query

Search for a single term:

```bash
python3 ~/.claude/skills/google-image-search/scripts/google_image_search.py \
  --query "neural interface wearable device" \
  --output-dir ./images \
  --num-results 5
```

### 2. Batch Processing

Process multiple queries from JSON config:

```bash
python3 ~/.claude/skills/google-image-search/scripts/google_image_search.py \
  --config image_queries.json \
  --output-dir ./images \
  --llm-select
```

### 3. Generate Config from Terms

Create JSON config from a list of terms using LLM:

```bash
python3 ~/.claude/skills/google-image-search/scripts/google_image_search.py \
  --generate-config \
  --terms "AlterEgo wearable" "sEMG electrodes" "BCI headset" \
  --output my_queries.json
```

### 4. Enrich Obsidian Note

Extract visual terms from note, find images, and insert below headings:

```bash
python3 ~/.claude/skills/google-image-search/scripts/google_image_search.py \
  --enrich-note ~/Brains/brain/Research/neural-interfaces.md
```

This mode:
1. Detects Obsidian vault and attachments folder
2. Uses LLM to extract visual-worthy terms from note
3. Searches for images for each term
4. Downloads best images to attachments folder
5. Inserts image embeds below relevant headings
6. Creates backup before modifying note

## Key Options

| Option | Description |
|--------|-------------|
| `--query TEXT` | Simple single query |
| `--config FILE` | JSON config for batch |
| `--generate-config` | Generate config from `--terms` |
| `--enrich-note FILE` | Enrich Obsidian note |
| `--output-dir DIR` | Where to save images |
| `--urls-only` | Return URLs only, no download |
| `--llm-select` | Use LLM to pick best image (default: on) |
| `--no-llm-select` | Disable LLM selection |
| `--num-results N` | Results per query (default: 5) |
| `--dry-run` | Show what would be done |

## JSON Config Format

Each entry supports:

```json
{
  "id": "unique-id",
  "heading": "Display Heading",
  "description": "Context for what image to find",
  "query": "Google search query",
  "numResults": 5,
  "selectionCriteria": "What makes a good image",
  "requiredTerms": ["must", "have"],
  "optionalTerms": ["bonus", "terms"],
  "excludeTerms": ["stock", "clipart"],
  "preferredHosts": ["official-site.com"],
  "selectionCount": 2
}
```

See `references/api_config_reference.md` for full documentation.

## Scoring System

Images are scored based on:
- **Required terms**: -80 if missing, +30 if all present
- **Optional terms**: +5 per match
- **Exclude terms**: -50 per match
- **Preferred hosts**: +25 if trusted, -5 if unknown
- **MIME type**: +5 for PNG/JPEG, -10 for GIF
- **Resolution**: +10 for high res, -10 for low res
- **File size**: -5 if very small

## LLM Selection

After scoring, LLM picks the best image from top candidates based on:
- Title and URL metadata
- Scoring reasons
- Selection criteria

The LLM evaluates authenticity, clarity, and relevance for technical audiences.

## Obsidian Integration

When in an Obsidian vault:
- Auto-detects vault root via `.obsidian` folder
- Uses configured attachments folder (default: `Attachments`)
- Generates Obsidian-style embeds: `![[image.png|alt text]]`
- Creates backup before modifying notes

## Script Files

| File | Purpose |
|------|---------|
| `google_image_search.py` | Main entry point |
| `api.py` | Google Custom Search API |
| `config.py` | Credentials and config handling |
| `download.py` | Image download with magic bytes |
| `evaluate.py` | Keyword-based scoring |
| `llm_select.py` | LLM selection and term extraction |
| `obsidian.py` | Vault detection and enrichment |
| `output.py` | Markdown output generation |