The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill creates a persistent data surface by ingesting untrusted user input (business descriptions, pain points, team size) and storing it in session memory (## Контекст бизнеса) for use by other skills.
Ingestion points: User responses to the five onboarding questions defined in reference/onboard-checklist.md.
Boundary markers: The data is stored in a specific Markdown block; however, the skill does not explicitly instruct the agent to ignore potentially malicious instructions embedded in these user-provided strings when they are read by subsequent skills.
Capability inventory: The skill can write to session memory and trigger external recipes/functions such as cash-flow-snapshot, crm-maintenance, and business-pulse.
Sanitization: No programmatic sanitization or escaping of user input is implemented; the skill relies on a manual "approval gate" where the user reviews the draft profile before it is saved.
[SAFE]: The skill incorporates robust "approval gates" (SKILL.md) that prevent the agent from silently recording data or connecting tools without the user's explicit consent and review. This manual confirmation step mitigates risks associated with data misuse.
[SAFE]: The skill is a localized version of established onboarding guidelines (derived from Anthropic's public knowledge-work-plugins) and does not exhibit signs of malicious intent, credential theft, or unauthorized remote code execution.

smb-onboard