scrask-bot

# Scrask Bot

## Overview

This skill activates when the user sends a screenshot via **Telegram**.
It uses vision AI to extract actionable information from the image, then:

- **High confidence (≥ 0.75):** Saves immediately and replies with a brief confirmation.
- **Low confidence (< 0.75):** Shows a structured preview in Telegram and asks for confirmation before saving.

**Provider behaviour (auto mode, default):**

| Step | What happens |
|---|---|
| 1 | Gemini 2.0 Flash parses the screenshot (fast, cheap) |
| 2 | If any item confidence < 0.60, Claude Opus reruns the parse |
| 3 | Whichever provider scores higher average confidence wins |
| 4 | Output includes `provider`, `fallback_triggered`, and confidence delta |

Set `vision_provider` to `"claude"` or `"gemini"` to lock a specific provider.

**Output destinations (AI-decided by content type):**

| Detected type | Destination |
|---|---|
| Event (has date+time, venue, or invite link) | Google Calendar |
| Reminder (deadline, due date, personal action) | Google Tasks (with due date) |
| Task (no date, pure action item) | Google Tasks (no due date) |

---

## Trigger Conditions

Activate when:
1. The user sends a message in Telegram that contains an **image attachment**
2. The image appears to be a **screenshot** — not a photo of a person, place, or physical object
3. No other skill has already claimed the image

Do not activate for:
- Photos of people, places, food, scenery
- Screenshots of code, errors, or UI bugs (leave for other skills)
- Images the user explicitly asks to edit, describe, or analyze for another purpose

---

## Step-by-Step Instructions

### Step 1: Acknowledge Immediately

Reply in Telegram right away so the user knows the skill is working:

> "📸 Got it — analyzing your screenshot..."

Do not make the user wait silently.

---

### Step 2: Run the Parser

```bash
python3 ~/.openclaw/skills/scrask-bot/scripts/scrask_bot.py \
  --image-path "<path-to-temp-image>" \
  --provider "$CONFIG_VISION...

# 🦞 Scrask Bot

**OpenClaw Skill** — Send a screenshot to Telegram. Scrask saves it to Google Calendar or Tasks automatically.

**Scrask** = Screenshot + Task

---

## What It Does

1. You take a screenshot on your phone (WhatsApp forward, email, social post, chat)
2. You send it to your OpenClaw bot on Telegram
3. Scrask parses it using vision AI
4. It saves it to the right place — no input needed from you

| Detected type | Destination |
|---|---|
| Event (date + time / venue / invite link) | Google Calendar |
| Reminder (deadline, due date) | Google Tasks (with due date) |
| Task (no date, action item) | Google Tasks |

High confidence (≥ 0.75) → saves silently, confirms in chat  
Low confidence → shows preview, asks before saving

---

## Provider Strategy (v3)

By default, Scrask uses **auto mode**: Gemini first, Claude fallback.

```
Screenshot arrives
      ↓
  Gemini 2.0 Flash (fast, cheap)
      ↓
  Any item confidence < 0.60?
  ├── No  → Done ✓
  └── Yes → Claude Opus reruns the parse
              ↓
          Claude avg confidence > Gemini + 0.05?
          ├── Yes → Use Claude result ✓
          └── No  → Gemini result was fine, keep it ✓
```

You can override this per-use with `--provider claude` or `--provider gemini`.

**What you get in the output:**

```json
{
  "provider": "claude",
  "fallback_triggered": true,
  "gemini_avg_confidence": 0.51,
  "claude_avg_confidence": 0.82,
  "confidence_gain": 0.31
}
```

---

## Installation

```bash
# 1. Copy to OpenClaw skills directory
cp -r scrask-bot ~/.openclaw/skills/

# 2. Install dependencies
pip install -r ~/.openclaw/skills/scrask-bot/scripts/requirements.txt

# 3. Set up Google credentials
# → Google Cloud Console → create service account
# → Enable Calendar API + Tasks API
# → Download JSON key → save as ~/.openclaw/google-creds.json
# → Share your Google Calendar with the service account email

# 4. Add to openclaw.json (see below)

# 5. Restart OpenClaw
openclaw restart
```

### openclaw.json co...