The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to clone and audit untrusted external repositories. This creates a significant surface for indirect prompt injection, where malicious instructions hidden in a competitor's README, source code, or documentation could influence agent behavior.
Ingestion points: Files (README.md, package.json, source code) from external repositories are cloned via git clone and read by the agent.
Boundary markers: The skill mandates strict line-numbered citations (nl -ba) and includes a 'Forbidden' patterns section to prevent the agent from hallucinating implementation details or following unsupported inferences.
Capability inventory: The skill uses network tools (git, gh), file system operations (mkdir), and includes references for potential code execution.
Sanitization: The skill relies on procedural constraints (citation rules and 'Preflight' checks) rather than automated content sanitization.
[REMOTE_CODE_EXECUTION]: In references/mechanism_evidence.md, the skill describes a workflow for 'executing competitor code' to trace behavior. While it instructs the agent to 'apply the current task's authorization boundary' and identify effects before execution, this pattern involves running untrusted external code.
[COMMAND_EXECUTION]: The skill uses several shell scripts and CLI tools to manage the analysis environment:
scripts/update-competitors.sh performs git clone, git fetch, and git pull operations on external URLs.
tests/test_reconcile_opportunities.py uses subprocess.run to execute a local Python script for data reconciliation.
The agent is instructed to use gh search repos for competitor discovery.
[DYNAMIC_EXECUTION]: The testing suite (tests/test_reconcile_opportunities.py) uses importlib.util and exec_module to dynamically load the scripts/reconcile_opportunities.py module at runtime for validation purposes.

competitors-analysis