The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill extracts and executes shell commands defined in the project's CLAUDE.md file under the 'Development Commands' section.
Evidence: The script in Step 0.6 parses {install-command} and {postinstall-command} from CLAUDE.md and executes them directly in the worktree environment.
Context: While intended for standard setup (e.g., bun install), this mechanism allows for the execution of arbitrary commands if the configuration file is malicious.
[COMMAND_EXECUTION]: The skill makes extensive use of system shell commands to manage the workspace and generate identifiers.
Evidence: Uses find, wc, xargs, and sed for workspace detection; shasum and date for ID generation; and various git worktree and git commit commands for state management.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection by processing external data and local project files to drive its logic.
Ingestion points: The skill reads content from CLAUDE.md and multiple files within the @docs/constitutions/current/ directory (e.g., architecture.md, patterns.md).
Boundary markers: Absent. The skill does not implement specific delimiters or 'ignore' instructions when reading these local configuration files.
Capability inventory: The skill has access to the bash tool, enabling file creation, directory manipulation, and execution of commands retrieved from the project context.
Sanitization: Absent. Content from the ingested files is interpolated into the workflow and used to determine which commands to run without verification.

writing-specs