Guide · Agent skill
How to use the Competition Agent Cloud agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
competition-agent-cloud is a specialist agent skill for capture-the-flag (CTF) security challenges that involve AI agents, prompt injection, MCP or toolchains, cloud, containers, CI/CD and supply chains. This guide explains what a skill is, how to install this one from its ZIP, how to ask for it in plain language, and what to watch for when running it. It is meant for players and researchers working inside a sandbox, and it only works properly after its parent skill is in use.
Competition Agent Cloudzhaoxuya520/reverse-skill · MIT Download the skill Quick start
- 1Download the ZIP. Download the competition-agent-cloud skill ZIP from this page.
- 2Unzip the skill folder. Unzip the file so the competition-agent-cloud folder contains SKILL.md, agents/openai.yaml and references/agent-cloud.md.
- 3Place it in the skills folder. Move the folder to ~/.claude/skills/ for personal use, or to the project's .claude/skills/ for one project.
- 4Or upload it in the Claude apps. In the Claude apps, go to Settings → Capabilities → Skills and upload the ZIP.
- 5Install the parent skill. Make sure ctf-sandbox-orchestrator is also installed, since this skill routes back to it.
- 6Start with the parent workflow. Begin your session with the parent orchestrator so sandbox assumptions are set before this skill is used.
- 7Ask in plain language. Describe your prompt-to-tool, cloud, CI/CD or supply-chain challenge and ask for one minimal chain to be traced.
What is an agent skill, and what does this one do?
An agent skill is a folder of instructions and reference files that a skills-compatible agent loads when a task matches its description. The agent reads the skill's SKILL.md and, when needed, the files it points to.
competition-agent-cloud is a downstream skill for the ctf-sandbox-orchestrator parent skill. Once the parent has established sandbox assumptions, node ownership and evidence priorities, this skill helps the agent trace how untrusted input can drive an AI agent's tools, and how deployment manifests, mounted secrets and build or release artifacts differ from what is actually running.
- Trace one minimal chain: untrusted input, then model-visible context, then a tool or deployment side effect.
- Separate what a prompt or manifest claims from what the runtime actually exposes.
- Compare checked-in manifests with live mounts, environment, logs and traffic.
- Trace provenance from source through dependency resolution, build and publish to the runtime consumer.
- Keep compact evidence such as prompts, tool arguments, manifests and hashes.
Install the competition-agent-cloud skill
Download the skill ZIP from this page and unzip it. The skill folder contains SKILL.md, agents/openai.yaml and references/agent-cloud.md.
You can install it in one of two ways. For a personal install, unzip it into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder. Alternatively, in the Claude apps, open Settings, go to Capabilities, then Skills, and upload the ZIP there.
- Personal: place the folder in ~/.claude/skills/
- Project only: place the folder in the project's .claude/skills/
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP
- Also install the parent skill, ctf-sandbox-orchestrator, because this skill routes back to it
competition-agent-cloud/
SKILL.md
agents/openai.yaml
references/agent-cloud.mdTrigger the skill with natural requests
You do not need a special command. Describe the task in plain language and name the kind of problem you are working on. The skill's description lists the topics it covers, so requests that match those topics are the ones most likely to load it.
The skill is designed to run only after ctf-sandbox-orchestrator has established sandbox assumptions. If the parent has not run yet, the skill tells the agent to return to the orchestrator first, so start a session with the parent workflow.
- Analyze how a prompt leads to a tool call in this challenge sandbox.
- Check whether the mounted secrets in the compose file match what the running container sees.
- Find where the deployed image differs from the one the CI pipeline built.
- Investigate possible retrieval poisoning in the document store used by the agent.
- Compare the runtime tool list with what the MCP manifest claims.
Examples of what the skill helps with
The skill focuses on evidence. For an agent challenge, it asks the agent to map instruction layers, retrieval layers, tool gates and side effects, then prove one minimal exploit chain before exploring variants. For a cloud or CI challenge, it asks the agent to keep build-time, deploy-time and runtime separate.
When a task belongs more narrowly to another competition topic, the skill points to a more specific sibling skill, such as prompt injection, supply chain, SSRF, container runtime or Kubernetes control plane work.
- Prompt injection through retrieved documents into a tool call
- Supply-chain drift from a dependency version to a published image
- Live container mounts that differ from a checked-in compose file
- Metadata-derived token paths in a sandboxed cloud scenario
Safety notes: scripts, network and scope
This skill runs no scripts and uses no network access, according to its listing. Its instructions are text that guide the agent's reasoning and file reading.
The skill is written for sandboxed competition challenges. It does not include authorization or scope checks. Do not point it at production systems or third-party infrastructure. Its instructions also say to reply in Simplified Chinese unless English is requested, so set that preference if you want English output.
- Use it only on sandboxed CTF challenges or lab environments you are allowed to test.
- It contains no authorization or scope checks, so you must confirm scope yourself.
- It does not run scripts and does not use the network.
- Replies default to Simplified Chinese unless you explicitly ask for English.
Tips for getting useful results
Give the agent the challenge files, manifests, prompts and logs up front, since the skill asks for compact evidence blocks rather than broad exploration. Ask for one minimal chain first and expand only after it is reproduced.
Watch for the common pitfalls named in the reference file: trusting a prompt string without runtime confirmation, and treating a checked-in manifest as proof of what was deployed.
- Start with the parent skill, then describe the single chain you want to test.
- Provide hashes, compose or Kubernetes fragments and logs where you have them.
- Ask for runtime confirmation before accepting a claim from a prompt or manifest.
What people use it for
- Tracing how untrusted retrieved text can influence an agent's tool arguments in a CTF sandbox
- Comparing a deployment manifest with live mounts and environment in a lab container setup
- Checking whether a CI-produced artifact matches the image that is running
- Reviewing an MCP or toolchain for claimed versus runtime-exposed capabilities
Frequently asked questions
Is the competition-agent-cloud skill free to use?
The source material for this page does not state a price or licence terms beyond the listing. The item is listed under the MIT licence, so check that licence text before reuse. The skill itself is a set of text files with no paid service attached.
What do I need before installing it?
You need a skills-compatible agent runtime and the parent skill, ctf-sandbox-orchestrator, installed. Sibling competition skills are also needed for full routing to more specific topics.
Where do I put the skill folder?
Unzip it into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder for one project. In the Claude apps you can instead upload the ZIP under Settings → Capabilities → Skills.
Why does the skill not work on its own?
Its instructions say to use it only after ctf-sandbox-orchestrator has established sandbox assumptions and routed the task here. If the parent has not run, the agent is told to return to the orchestrator first.
Does this skill run scripts or access the internet?
According to its listing, it runs no scripts and uses no network. It is a set of instructions and reference material for the agent.
How does it compare with other security skills?
It is a downstream specialist for agent, toolchain, cloud and supply-chain CTF work. For narrower topics such as prompt injection, supply chain, SSRF or container escape, the skill itself recommends the more specific sibling skills.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
competition-agent-cloud suits students and researchers who already work inside a sandboxed CTF setup and want help tracing prompt-to-tool flows and deployment drift. It is not suited to production or third-party systems, to use without the parent orchestrator skill, or to general agent, cloud or CI/CD building tasks, where a different resource would serve better.
Download the skill