The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions frequently use npx to download and execute the claude-flow package (and its alpha versions) from the public NPM registry. It also points users to external documentation and issue trackers on GitHub.
[COMMAND_EXECUTION]: The methodology is heavily based on executing shell commands via the claude-flow CLI to run various development modes such as coder, tdd, and researcher.
[DYNAMIC_EXECUTION]: The coder and orchestrator modes involve autonomous code generation and task delegation where natural language descriptions are translated into file operations and script executions by the underlying framework.
[INDIRECT_PROMPT_INJECTION]: The skill has a broad attack surface for indirect injection due to its data ingestion capabilities.
Ingestion points: The researcher mode uses WebSearch and WebFetch to gather external data, while the reviewer and analyzer modes process code from local files and GitHub repositories.
Boundary markers: There are no specific delimiters or instruction-bypass warnings mentioned for handling data retrieved from external sources.
Capability inventory: The skill utilizes a powerful set of capabilities including file writing (TodoWrite), network access, and code execution through the multi-agent orchestration framework.
Sanitization: The methodology documentation does not describe explicit sanitization or validation steps for content ingested from web searches or repository analysis.

sparc-methodology