How to use the AI Image Realism agent skill

What an agent skill is and what this one does

An agent skill is a folder containing a SKILL.md file and sometimes scripts that an AI agent reads and follows when a task matches the skill's description. This entry for ai-image-realism is a pointer to an installable package, not the working code itself.

The package is aimed at making existing AI portraits, product images and marketing visuals look more natural. The work runs on the hosted Beatra service at mcp.beatra.ai, and each paid render is billed.

• Works from inside Claude Code, Codex or OpenClaw, per the skill's overview.
• Targets plastic skin, malformed hands or faces, repeated textures and inconsistent lighting.
• Processing happens on a third-party paid service, so local-only editing is not what this skill does.

What you need before installing

Check these requirements first. The install flow fetches code from GitHub and opens a browser sign-in, so you need both network access and permission to make those changes on your machine.

• Python 3.10 or newer. The package description says it uses the standard library only.
• Git, used to fetch the pinned package.
• A Beatra account with paid credits, and the ability to complete a browser sign-in for authorization.
• Network access to github.com and mcp.beatra.ai.
• The package runs scripts and uses the network.

Install the skill from the ZIP or the Claude apps

Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ folder to limit it to that project.

Alternatively, in the Claude apps, open Settings, then Capabilities, then Skills, and upload the package there.

The skill's own instructions use a stricter, reviewed install in four steps: fetch the pinned tree, inspect it, copy it in, then authorize. Each step needs its own approval and stops if the expected IDs do not match.

• Step 1 fetches the package into a temporary review directory and checks the tree ID.
• Step 2 inspects the files, including scripts/authorize.py, scripts/mcp_client.py and scripts/uninstall.py.
• Step 3 copies the reviewed folder to ~/.claude/skills/ai-image-realism and refuses to overwrite an existing folder.
• Step 4 runs scripts/authorize.py, which opens a browser sign-in and stores a bearer token.

Run the free connection check

Before any paid work, run the non-billable verify command. It checks the connection and does not charge credits.

Start a new agent session afterwards if your host only discovers skills at startup.

Trigger the skill with a natural request and approve each charge

Ask for a specific retouch on an image you have supplied, and say that paid processing is acceptable. For example: "Retouch this product photo so the skin and hands look natural, and show me the cost before doing anything."

The agent is designed to go through a fixed sequence. It first uses free discovery to read current model cards and credit estimates, then shows a cost card for exactly one paid call listing the tool, model, estimate and files to be uploaded. It waits for explicit approval; a general instruction to proceed does not count.

After the render, the agent reports the artifact and the net charged credits, and asks again before any further paid stage.

• Good request: name the image, the flaw to fix, and that you accept the charge.
• Planning or "just do it" is not approval; the agent should wait for a yes to the cost card.
• Only files you supplied and approved for this task are uploaded.

Security and safety notes

The catalog entry itself contains no client code. The linked package bundles Python scripts that make network calls and store a sign-in token, so treat the install as a real software installation.

The package has no self-updater and does not download code or change its own files. Updating means repeating the review steps for a new commit and replacing the folder only after approval.

Uninstalling runs scripts/uninstall.py. Use --dry-run first to see the decision. The script can revoke the token and remove files under ~/.beatra; it never deletes recursively.

• Read every file in the pinned tree yourself before activation; grep output alone is not a source review.
• Do not install if you are uncomfortable with browser sign-in and a stored bearer token.
• Images leave your machine for a third-party service.