The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to manage a full software development lifecycle, which includes executing build scripts (./build_arkweb.sh) and error analysis scripts (scripts/analyze_build_error.sh) using the Bash tool. This involves high-privilege shell access and the execution of generated code.
[DATA_EXFILTRATION]: The environment pre-check process defined in reference/env-check.md reads and modifies the sensitive Claude Code configuration file located at ~/.claude/settings.json. It specifically injects the CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 environment variable. While documented as a setup requirement, modifying tool-wide security and environment configurations is a sensitive operation that alters the execution environment of the agent.
[PROMPT_INJECTION]: The skill ingests external requirement documents and design specifications which are then processed by a chain of agents (Requirements Analysis, Coder, Reviewer). This multi-step pipeline is susceptible to indirect prompt injection, where malicious instructions embedded in the input documentation could manipulate the coding agent to perform unauthorized actions or inject vulnerabilities into the project.
Ingestion points: The SKILL.md orchestrator and requirements-analysis.md agent take a document path provided by the user and read its full content into the context.
Boundary markers: The instructions include high-level mandates to follow the design documents strictly, but they lack technical delimiters or explicit "ignore embedded instructions" warnings for the interpolated content.
Capability inventory: The arkweb-coder agent is equipped with powerful tools including Bash, Write, and Edit, enabling it to modify the source code and execute commands based on the analyzed requirements.
Sanitization: There is no evidence of programmatic sanitization or filtering of the external content before it is processed by the agent chain.

arkweb-code-gen