I got a flatbed scanner. My immediate goal: scan handwritten notebook pages and make the text available to my AI coding agent without retyping anything.

Seemed simple. It was not. Here’s what I learned.

The setup

Problem 1 — the scanner didn’t work at all

Plugged it in via a USB hub. Windows detected it (Get-PnpDevice showed Status: OK). NAPS2 showed the device. Everything looked right.

But every scan attempt returned 0 pages.

Root cause: the LiDE 400 is bus-powered — it draws all its power from USB. My hub wasn’t delivering enough current. Moving to a direct port fixed it immediately.

If your flatbed scanner detects but doesn’t scan: try direct USB first before debugging anything else.

Problem 2 — Windows 11 24H2 broke ESCL USB scanning

The LiDE 400 supports two scan protocols: classic WIA and eSCL (the driverless network/USB protocol). I tried ESCL first since it doesn’t require manufacturer software.

It didn’t work. After some digging: Microsoft acknowledged a bug in Windows 11 24H2 where USB scanners using eSCL fail to enumerate correctly. The fix shipped in KB5048667 (December 2024). If you’re on a fresh 24H2 install without updates, eSCL-over-USB is broken.

WIA works fine even without the fix. So: use WIA.

The scan command

NAPS2’s CLI handles everything:

NAPS2 console \
  --driver wia \
  --device "CanoScan LiDE 400" \
  --source glass \
  --dpi 300 \
  --bitdepth color \
  -o scan.jpg \
  -f -v

Wrapped this in an OMP extension:

const child = spawn(NAPS2, [
  "--driver", "wia",
  "--device", DEVICE,
  "--source", "glass",
  "--dpi", "300",
  "--bitdepth", "color",
  "-o", out,
  "-f", "-v",
], { stdio: "pipe", windowsHide: true });

stdio: "pipe" is critical — same lesson from the screenshot extension. Without it the child process inherits OMP’s PTY and gets killed.

Result: /scan captures the page and attaches @C:/tmp/scans/scan-<ts>.jpg to the editor. Scan and ask in one step.

Problem 3 — Tesseract can’t read cursive

NAPS2 has built-in OCR via Tesseract:

NAPS2 console ... --enableocr --ocrlang eng -o scan.pdf

The PDF is created. The text layer is garbage:

mee
oat af Reh aide
"aaeatd Pg g&
FR sav ® SEBO ey

That’s from a page of legible (to humans) cursive English. Tesseract is trained on printed text. It’s not even close on handwriting.

This is a known, hard limitation. Tesseract doesn’t do handwriting.

The fix — vision LLM

Instead of OCR, send the image to a vision-capable LLM. The same Anthropic Claude API that your agent uses for reasoning handles image-to-text natively:

import base64, json, urllib.request

with open(image_path, "rb") as f:
    b64 = base64.b64encode(f.read()).decode()

payload = {
    "model": "claude-sonnet-4-5",
    "max_tokens": 2048,
    "messages": [{
        "role": "user",
        "content": [
            {
                "type": "image",
                "source": {"type": "base64", "media_type": "image/jpeg", "data": b64}
            },
            {
                "type": "text",
                "text": (
                    "Transcribe all handwritten text from this scanned notebook page. "
                    "It may be upside-down or rotated — correct mentally. "
                    "Preserve structure, bullet points, and indentation. "
                    "Mark uncertain words with [?]."
                )
            }
        ]
    }]
}

req = urllib.request.Request(
    "https://api.anthropic.com/v1/messages",
    data=json.dumps(payload).encode(),
    headers={
        "Content-Type": "application/json",
        "x-api-key": os.environ["ANTHROPIC_API_KEY"],
        "anthropic-version": "2023-06-01",
    }
)

with urllib.request.urlopen(req, timeout=30) as resp:
    result = json.loads(resp.read())

print(result["content"][0]["text"])

Output on the same page that Tesseract mangled:

**Meeting notes — Q3 planning**

- **Action items:**
  - ① Follow up with design team on spec
  - share draft by Friday
  - schedule review [?]

- Blocked on infra
- Need sign-off from Alice
- deadline: end of month

**Budget** → confirm with finance
- carry over from Q2

- risks: timeline slipping
- dependencies not resolved
- escalate if no response by Weds

Legible. Structure preserved. Uncertain words marked.

Model benchmark — what actually works for handwriting

I tested several models via an OpenAI-compatible gateway. Results:

Tier Model Notes
🥇 Best Claude Sonnet / Claude Opus (Anthropic API) Accurate, preserves structure, handles upside-down
🥉 Partial GPT-4o (Azure OpenAI) Works but less accurate on cursive
❌ Broken Gemini 2.5 Flash/Pro (via Vertex gateway) Image bytes don’t reach the model — returns text-only response

Two key findings:

  1. Don’t resize before sending. I tried resizing to 1024px to reduce payload size. GPT-4o’s accuracy dropped significantly. Send full resolution — Claude handles it fine.
  2. Gemini image routing is gateway-dependent. Gemini works fine via Google’s own API. Via an OpenAI-compatible proxy, the image payload may be silently dropped. Check your gateway’s multimodal support before assuming it works.

Two commands, you choose

I ended up with two slash commands in the OMP extension:

/scan — fast, no API cost, inline vision:

  1. NAPS2 scans → scan-<ts>.jpg
  2. @path attached to editor
  3. You type your question, I read the image

/scan-ocr — transcription saved to disk:

  1. NAPS2 scans → scan-<ts>.jpg
  2. Python calls vision API → scan-<ts>.txt
  3. Both @image and @txt attached

/scan is right for interactive Q&A — the vision model in the chat reads the image directly, no extra API call. /scan-ocr is right when you want a persistent text file for search or reuse.

Context window note: a 300 DPI scan (~2550×3300px) costs ~1,800 tokens as an image. The equivalent text transcript is ~300 tokens. If you’re referencing the same scan across many turns, the transcript wins.

One placement quirk

Scan the page face-down with the top edge at the arrow mark on the glass. If you place it the other way, the image comes out upside-down. Claude reads it correctly anyway (it notes the orientation and corrects mentally), but it’s cleaner to get it right.

Source