ChatMask Skill

When the user sends chat screenshots and asks to pixelate or hide identity
elements, follow the steps below in order: Setup → Workflow.

No external API key is required. This skill uses your existing AI
capabilities to locate regions in the image, then delegates only the image
manipulation to the local Python script.

Setup (run once, idempotent)

Run this block before the first job. Every step is guarded so re-running is safe.

CODEBLOCK0

What this installs: Pillow (image processing) and python-dotenv (.env loader
for standalone use only). The requests package is not installed by the
skill — it is only needed for standalone/non-skill mode and lives in
requirements-standalone.txt. No network calls are made by the script at runtime.

Network behavior: The only outbound calls in Setup are git clone (one-time)
and pip install (one-time). During Workflow, process.py makes no network
calls when --bbox-json is supplied — all processing is local.

Workflow

Read Element Selection and Option Configuration to translate natural-language
requests into the correct flags before running.

1. Prepare directories

CODEBLOCK1

2. Process each image individually

Each image must be analysed and pixelated separately. Different screenshots
have different element positions — passing one image's bounding boxes to another
would leave sensitive regions unredacted. Repeat the following block for every
image the user sent.

Why per-image? --bbox-json is scoped to a single image. process.py
enforces this: it exits with an error if more than one image is present in the
input directory when --bbox-json is used.

For each image <filename.png>:

2a. Copy the image into its own input directory

CODEBLOCK2

2b. Analyse the image with your vision capabilities

Use the prompt below on $IN_DIR/$IMG_FILE and capture the JSON output.
The schema uses normalized coordinates (0-1000) where (0,0) is top-left
and (1000,1000) is bottom-right, in y_min, x_min, y_max, x_max order.

CODEBLOCK3

Omit keys for elements the user did not request (see Element Selection below).

2c. Run pixelation for this image

Pass the JSON from 2b via --bbox-json. No API key is read or written.

CODEBLOCK4

Repeat steps 2a–2c for every image before proceeding.

3. Return results to user

CODEBLOCK5

Attach or share all processed images from $OUT_DIR/.

Element Selection

Translate the user's intent to --elements. Default (no flag) pixelates all three.

When not all three elements are requested, omit the unused keys from the
bounding-box JSON prompt in step 2 to reduce noise.

Element definitions:

• - chatname: Title text in the top navigation bar (group name, contact name, channel title)
• profilepic: Circular/rounded avatar images next to message bubbles
• display_name: Text username/nickname labels next to or above message bubbles

Option Configuration

Full Example (copy-paste ready)

Two images processed, each with its own per-image bounding-box analysis:

CODEBLOCK6

More examples

CODEBLOCK7

Troubleshooting