imagegen

Files (1)

• SKILL.md

SKILL.md (first lines)

# imagegen

Generate design concept images using image generation models. The workflow: produce visual candidates quickly, review with the navigator, then use the best results as references for Figma vector work — fast-tracking ideation before the slower modeling step.

## Pre-flight: credential discovery

The skill needs a Gemini API key. Resolve it in this order (first match wins):

1. **Environment variable:** check if `GEMINI_API_KEY` is set in the current shell environment.
2. **First-use prompt:** if the environment variable is not set, tell the user:
   > "I need a `GEMINI_API_KEY` to generate images. You can get one at https://aistudio.google.com/apikey — paste it here, or add it to your `.envrc.local` and reload direnv."

   If the user provides the key inline, use it for the current session. Suggest they persist it:
   > "To avoid this prompt next time, add this to your `.envrc.local`:"
   > ```
   > export GEMINI_API_KEY="your-key-here"
   > ```
   > "If you use 1Password, you can reference it as:"
   > ```
   > export GEMINI_API_KEY=$(op read "op://path/to/your/key")
   > ```

Never store the key in committed files, skill files, or memory. The key lives in the user's environment only.

## Pre-flight: output directory

Before generating images, verify the output directory setup:

1. Create `.imagegen-output/` in the project root if it doesn't exist.
2. Check if `.imagegen-output/` appears in `.gitignore`. If it does not:
   - Warn the user: "`.imagegen-output/` is not in `.gitignore` — generated images could be accidentally committed."
   - Offer to append it: "Shall I add it? I'll include a comment explaining what it's for."
   - If approved, append to `.gitignore`:
     ```
     # imagegen skill output (generated design concept images)
     .imagegen-output/
     ```

## Model selection

Pick the model based on the task. Start cheap/fast and upgrade only if quality falls short.

| Model | Speed | Quality | Best for |
|-------|-------|---------|----------|
| `gemini-2.5-flash-image` | Fast | Good | **Default.** Quick ideation, first-pass exploration, mood boards |
| `gemini-3.1-flash-image` | Medium | Higher | Refined concepts, detailed scenes, when 2.5 output lacks detail |
| `gemini-3.1-flash-lite-image` | Fastest | Lower | High-volume batches, throwaway sketches, rapid A/B comparisons |
| `gemini-3-pro-image` | Slow | Best | Complex multi-element compositions, high-res output, final reference images |

**Decision flow:**
1. Start with `gemini-2.5-flash-image` — it handles most ideation work.
2. If the output lacks detail or misses compositional nuance, upgrade to `gemini-3.1-flash-image`.
3. If you need high-res (2K/4K) or complex multi-object scenes, use `gemini-3-pro-image`.
4. If you're generating 10+ throwaway variants to narrow a direction, use `gemini-3.1-flash-lite-image`.

The user can override at any time: "use the pro model for this one" or "switch to lite, I just want quick sketches."

## API reference

**Endpoint:**
```
POST https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent?key={GEMINI_API_KEY}
```

**Request body:**
```json
{
  "contents": [
    {
      "parts": [
        { "text": "your prompt here" }
      ]
    }
  ],
  "generationConfig": {
    "responseModalities": ["IMAGE", "TEXT"]
  }
}
```

**Response structure:**
```json
{
  "candidates": [
    {
      "content": {
        "parts": [
          {
            "inlineData": {
              "mimeType": "image/png",
              "data": "<base64-encoded-image-data>"
            }
          },
          {
            "text": "optional text description from the model"
          }
        ]
      }
    }
  ]
}
```

**Authentication:** query parameter `?key=` (or `x-goog-api-key` header).

**Constraints:**
- Default output: 1024x1024 px
- Free tier: ~500 images/day
- Cost: ~$0.039/image at 1024px
- All outputs carry a SynthID watermark

### Execution

Use whatever HTTP tool is available — `curl` in the shell, a built-in HTTP fetch tool, or similar. The agent decides the mechanism. Example with curl:

```bash
curl -s -X POST \
  "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-image:generateContent?key=${GEMINI_API_KEY}" \
  -H "Content-Type: application/json" \

From asabina-de/kb, CC-BY-4.0.

More from asabina-de/kb