The Agent Skills Directory

[SAFE]: The skill follows a human-in-the-loop security model. It explicitly mandates 'approval gates' for every step, requiring the user to review side-by-side diffs before any changes are applied. It also includes hard constraints against deleting records or closing deals automatically.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external data from a CRM (deals, contacts, activities). While this is an ingestion surface for untrusted data, the risk is mitigated by the requirement for manual user confirmation of all proposed changes, which prevents the agent from performing unintended actions autonomously based on the data content.
[COMMAND_EXECUTION]: The Bash tool is listed in the allowed-tools frontmatter, which restricts the skill's environment rather than granting new powers. The instructions themselves do not contain any shell commands, script generation, or system-level modifications.
[PROMPT_INJECTION]: The instruction to "skip the intent determination step" (пропусти шаг определения намерения) is a functional directive to streamline the workflow for a specific command rather than an attempt to bypass safety filters or ignore prior system instructions.

crm-cleanup