How to use the CTF Web Runtime Analyst agent skill

What an agent skill is and what this one does

An agent skill is a folder of instructions and reference files that an AI agent loads when a task matches the skill's description. This one is an internal helper for the larger ctf-sandbox-orchestrator workflow. It is written for CTF challenges involving web, API, server-side rendering, frontend, queue-backed apps, and routing.

Once active, the skill directs the agent to follow a fixed order. It first assumes the presented hosts and routes belong to the sandbox, then inspects entry HTML, boot scripts, runtime config, and route registration. It captures one real request flow end-to-end before making broad claims from source code, checks browser persistence and backend state together, and re-runs the smallest flow with one variable changed.

• Maps active hosts, paths, proxies, containers, and workers
• Inspects cookies, localStorage, sessionStorage, IndexedDB, Cache Storage, and service workers
• Records exact host, path, query, headers, cookies, and body for decisive requests
• Traces middleware, auth and session boundaries, uploads, exports, and background jobs once one path is proven

Install the competition-web-runtime skill

Download the skill ZIP from this page and unzip it. For a personal setup, place the unzipped folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder.

If you use the Claude apps instead, you can upload the skill under Settings → Capabilities → Skills. Either way, the agent needs the skill files in place before it can load them.

The package includes agents/openai.yaml, SKILL.md, and two reference files under references/: cookie-hmac-key-reuse-auth-bypass.md and routing-runtime.md.

Set up the prerequisite parent skill first

This skill is not meant to run alone. It says it should only be used after $ctf-sandbox-orchestrator is already active and has established sandbox assumptions, node ownership, and evidence priorities. If that has not happened, the skill tells the agent to return to the parent skill first.

The SKILL.md also points to several sibling skills for specialized tasks. Examples include competition-browser-persistence for cookies and storage, competition-runtime-routing for Host headers and vhost routing, and competition-jwt-claim-confusion for JWT issues. These sibling skills are referenced but not included in this package.

• Required: ctf-sandbox-orchestrator (parent skill)
• Referenced, not included: competition-browser-persistence, competition-runtime-routing, and other specialized sibling skills

Trigger the skill with natural requests

The skill's description lists the kinds of requests that should activate it. You do not need special syntax. Describe the web task in plain language, and make sure the parent skill has already set up the sandbox context.

Requests that match the description include asking the agent to inspect a site or API, follow real browser requests, debug auth or session flow, trace uploads or workers, find hidden routes, or explain why frontend and backend behavior diverge under sandbox-internal routing.

• Inspect this sandbox site and tell me how the login request is sent.
• Follow the real browser requests when I click checkout and show me the order.
• Debug why the session cookie is accepted on the API but not on the admin page.
• Trace what happens to the file after it is uploaded, including any worker that picks it up.
• Find hidden routes the frontend never links to in this challenge.

Work through a challenge: what a good session looks like

The skill asks the agent to prove one path before expanding. A typical session starts with the entry page and its scripts, then records the decisive requests and responses, and only then looks at middleware, handlers, or background jobs.

The skill's reference material also describes a known pattern in which a server signs a cookie using a key that is exposed in the URL, and the backend trusts claims inside the cookie payload. The reference gives the agent a checklist for confirming whether that pattern applies, but the actual result depends on the challenge.

• Keep only the exact requests and responses that change the conclusion
• Record route names, storage keys, queue names, and worker names that appear in the active flow
• Treat UI gating as a hint, not proof that the backend enforces it

Safety, scope, and network use

The skill makes network requests, because its job is to capture and replay real browser and API traffic. It does not run scripts. The skill assumes every host it sees belongs to the sandbox, and it does not check authorization on its own.

Do not point it at live production systems or third-party sites you do not own or have permission to test. Only use it on the sandbox hosts that your challenge or team has provided.

• Runs scripts: no
• Uses network: yes
• No built-in authorization check; scope control is your responsibility

Tips for better results

Give the agent the decisive context up front: which host is in scope, what the goal is, and which step is failing. The skill is written to work from evidence, so a captured request is more useful than a description.

The SKILL.md says the agent replies in Simplified Chinese unless you explicitly ask for English. Say so in your request if you want English output.

• State the goal and the sandbox host before asking for analysis
• Share the failing request and response if you already have them
• Change only one variable per test run to keep results clear
• Ask for English output explicitly, since the default reply language is Simplified Chinese