The Agent Skills Directory

[COMMAND_EXECUTION]: The skill modifies the agent's security configuration by using jq to read the global ~/.claude/settings.json file and merge its permissions.allow and permissions.deny arrays into the local .claude/settings.local.json. This allows the project to inherit broad permissions, effectively bypassing local restriction defaults.
[COMMAND_EXECUTION]: The skill registers a PostToolUse hook (.claude/hooks/copy-paper-pdf.sh) in the agent's local settings. This causes a bash script to execute automatically whenever the Bash tool is used, providing a persistent mechanism for background command execution.
[COMMAND_EXECUTION]: The skill generates a .latexmkrc configuration file that contains a Perl system() call. This executes shell commands (cp) during the LaTeX compilation process, which is a form of dynamic command execution within a configuration file.
[COMMAND_EXECUTION]: The skill creates a run_all.sh script that executes Python, R, and Stata scripts via shell commands (uv run, Rscript, stata-mp). While functional, this provides a pre-configured vector for executing code that may be modified by other agents or users.
[PROMPT_INJECTION]: Phase 1.5 involves scanning existing files in a directory and reading their content to "absorb" it into the new project's metadata and configuration. This creates a surface for indirect prompt injection, as malicious instructions embedded in existing documents could be processed and acted upon by the agent during the initialization phase.
Ingestion points: Reads .tex, .md, and .bib files from the target directory during the auto-detection and reorganization phases (Phase 1.0 and 1.5).
Boundary markers: None identified for the scan of existing user files.
Capability inventory: The skill has extensive bash execution capabilities (mkdir, touch, chmod, git, jq, ln, rsync) and file write access.
Sanitization: No evidence of sanitization or filtering of content read from existing files before it is used to seed new configuration files or project metadata.
[EXTERNAL_DOWNLOADS]: The skill performs several external operations, including creating private GitHub repositories via the gh CLI, rsyncing templates from a local Task Management directory, and querying the Elsevier Serial Title API for journal rankings (which requires an API key).

init-project-research