Screenshot Debugger - Nito Plugin

Goal. Sometimes the fastest way to explain a problem is a picture: a stack-trace screenshot, a red build log, a misrendered UI, an architecture diagram. This recipe attaches an image to a vision-capable model through Nito and asks about it, from the terminal or from inside Claude Code or Codex, with a batch loop for a folder of screenshots. When to reach for it. A crash you captured as a screenshot, a layout bug that is easier to show than describe, a diagram you want explained or turned into notes, or a folder of QA screenshots to triage. For a document (PDF, DOCX), use Document Q&A Pipeline instead; images and documents take different paths.

Images Need a Vision Model

Documents work with any model because Nito extracts their text first. Images do not: the model itself has to accept image input, so you must pick a vision-capable model. Nito refuses an image on a non-vision model with a clear message rather than guessing. Supported image formats are PNG, JPEG, and WebP. See Image input. Find vision models with nito agent models (look for image support). qwen/qwen3-vl-235b-a22b-instruct (open-weight, Private) and anthropic/claude-haiku-4-5 (frontier, Anonymous) both accept images; pick by the privacy level you want.

Prerequisites

Nito installed and signed in. See Nito CLI.
A vision-capable model (see above).
Screenshots in PNG, JPEG, or WebP.

Step 1: Ask about one screenshot

From the terminal, attach the image with --file and name a vision model:

# Explain an error captured as a screenshot
nito ask --model qwen/qwen3-vl-235b-a22b-instruct --file ./error.png \
  "What is causing this error, and how do I fix it?"

Step 2: Read the result

Nito · Hosted · anthropic/claude-haiku-4-5 · Anonymous · Prompt with 1 image

The traceback shows a KeyError on "user_id" in serializers.py line 88. The
request payload does not include user_id, so `data["user_id"]` raises. Guard it
with data.get("user_id"), or validate the field before access.

The label confirms an image was attached (Prompt with 1 image) and which model and privacy level answered.

Inside a Chat

The same works with the plugin, using an image path in your workspace:

Claude Code
Codex

/nito:ask --model qwen/qwen3-vl-235b-a22b-instruct --file ./ui-bug.png Why is this button overlapping the header?

$nito:ask --model qwen/qwen3-vl-235b-a22b-instruct --file ./ui-bug.png Why is this button overlapping the header?

Triage a Folder of Screenshots

To triage many images at once, for example a directory of failing-test screenshots, loop over them and collect a one-line diagnosis per file:

#!/usr/bin/env bash
# triage-shots.sh [folder]
set -euo pipefail

DIR="${1:-./screenshots}"
MODEL="anthropic/claude-haiku-4-5"

shopt -s nullglob
files=("$DIR"/*.png "$DIR"/*.jpg)
[ ${#files[@]} -eq 0 ] && { echo "No images in $DIR" >&2; exit 0; }

for img in "${files[@]}"; do
  echo "=== $(basename "$img") ==="
  nito ask --model "$MODEL" --file "$img" --quiet \
    "In one sentence, what does this screenshot show and what is wrong, if anything?"
done

./triage-shots.sh ./failing-tests

Variations

Use a frontier vision model instead. The examples above run at the Private level. If you want a frontier model on the image and the screenshot is not sensitive, anthropic/claude-haiku-4-5 accepts images at the Anonymous level:

nito ask --model anthropic/claude-haiku-4-5 --file ./mockup.png \
  "Describe this layout."

Turn a diagram into notes. Point it at an architecture diagram or a whiteboard photo:

nito ask --model qwen/qwen3-vl-235b-a22b-instruct --file ./architecture.png \
  "List the components in this diagram and how data flows between them."

Review a UI mockup. Ask for an accessibility or UX critique:

nito ask --model qwen/qwen3-vl-235b-a22b-instruct --file ./mockup.png \
  "Give three accessibility issues in this design and how to fix each."

Troubleshooting

What you see	Why	Fix
”cannot accept image attachments; choose an vision-capable model”	The model does not do vision	Pick a vision model (check `nito agent models`)
An unsupported-format error	The file is not PNG, JPEG, or WebP	Convert the image, or use a document flow for PDFs
”file not found”	Wrong path	Use a path relative to where you launched the command
The model misreads small text	The screenshot is low resolution	Crop to the relevant area and re-capture at higher resolution

Where to Go Next

Image input

How image attachments work and which models accept them.

Document Q&A Pipeline

The document path, for PDFs and office files.

Choosing a Model

Finding vision-capable models and their privacy levels.

Scripting and Automation

Output modes and looping for batch triage.

​Images Need a Vision Model

​Prerequisites

​Step 1: Ask about one screenshot

​Step 2: Read the result

​Inside a Chat

​Triage a Folder of Screenshots

​Variations

​Troubleshooting

​Where to Go Next