Case 03484

Skill knownasnaffy/prompthound/dataset/case_03484

A fast, offline static risk analysis CLI for AI agent skill files. Detects malicious instructions, steganographic payloads, and dangerous capability chains.

Install

npx -y skills add knownasnaffy/prompthound --skill case_03484

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Generate and edit images with a bundled Gemini native image-generation CLI. Use when the user asks Codex to create images with Gemini, use Gemini image generation through a local bash CLI, run or debug Gemini image curl requests, choose Gemini image models/sizes/aspect ratios, inspect raw image-generation responses, or troubleshoot slow, timed-out, or failed Gemini image requests.

SKILL.md

3.1 KB, 644 tokens by cl100k_base, as published. Nobody here has run it

Gemini Image CLI

Use ./scripts/gemini-image.sh for Gemini native image generation. Prefer this bundled script over writing one-off curl commands.

Workflow

Run ./scripts/gemini-image.sh with the user's prompt and any requested options.
Do not ask which endpoint to use for ordinary requests. The script auto-selects the provider: local Gemini-compatible proxy first, then Google fallback.
Keep default settings for ordinary single-image generation: gemini-3.1-flash-image-preview, size 512, aspect 16:9.
Use gemini-2.5-flash-image when latency matters more than latest image quality.
Use gemini-3-pro-image-preview when the user needs stronger instruction following, text rendering, or professional-quality output.
Confirm before multi-model batches, many retries, or other repeated calls that may consume extra quota.
Read references/behavior.md only when explaining provider/security tradeoffs, choosing non-default models, configuring a local Gemini-compatible proxy, troubleshooting slow or failed requests, or modifying the CLI.

Common Commands

Generate one image:

./scripts/gemini-image.sh "A cute orange kitten sitting on a soft blanket"

Generate with an explicit output path or prefix. The script chooses the final extension from the returned image MIME type:

./scripts/gemini-image.sh "画两只小猫在打闹" --output ./out/kittens.png

Use a faster model:

./scripts/gemini-image.sh "画两只小猫在打闹" --model gemini-2.5-flash-image

Force Google official endpoint:

./scripts/gemini-image.sh "画两只小猫在打闹" --provider google

Force local proxy endpoint:

./scripts/gemini-image.sh "画两只小猫在打闹" --provider local

Use a larger output size:

./scripts/gemini-image.sh "A cinematic poster of two kittens" --size 1K --aspect 16:9

Use an input image for image-guided generation or editing:

./scripts/gemini-image.sh "Turn this cat photo into a watercolor illustration" --image cat.jpg

Output Contract

The script prints human-readable logs to stderr and machine-readable results to stdout.

Successful stdout lines:

image=<path>
raw_json=<path>
text=<path>
duration_seconds=<seconds>

text= appears only when --with-text is enabled.

Safety

Do not expose full Google Gemini API keys in conversation or source files. Prefer the local proxy mode when the runtime should not have access to the real Google key.

The script masks keys in curl logs and redacts input-image base64 from printed request bodies.

Do not enable retries automatically for ambiguous multi-request tasks. Retries can submit additional generation requests and may incur additional cost.

What ships with it: 2 files

33.3 KB alongside SKILL.md, 1 of them executable

references/

behavior.md4.4 KB

scripts/

gemini-image.shruns28.9 KB