The Agent Skills Directory

[COMMAND_EXECUTION]: The agent uses the claudeCode harness to execute shell commands, specifically cargo-ai storage query execute for CRM data retrieval and standard Git commands for repository management. These commands are strictly constrained by the system prompt, which forbids improvisation.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from external CRM systems, which creates a potential surface for indirect injection if deal records contain malicious instructions. The risk is mitigated by using fixed SQL queries and explicit prompt instructions that require the agent to use specific counts and denominators.
Ingestion points: CRM models (crm_deals, crm_accounts, crm_contacts) defined in the infra/models/ directory.
Boundary markers: The system prompt in infra/agents/win-loss-analyst.prompt.ts contains strict instructions to 'do not improvise' and 'never run a query that is not in step 1'.
Capability inventory: Shell execution via the coding harness, database querying, GitHub pull request creation, and Slack messaging.
Sanitization: The audit relies on structured SQL extraction of specific fields rather than unstructured text analysis.
[DATA_EXFILTRATION]: The skill extracts information from CRM databases and transmits summaries to Slack and GitHub. Security is maintained by locking the Slack channelId to a specific internal channel and using a human-gated pull request workflow for all repository updates.
[EXTERNAL_DOWNLOADS]: The documentation instructs users to install standard vendor-provided tools, including @cargo-ai/cli and @cargo-ai/cdk, which are necessary for the skill's deployment and verification scripts.

win-loss-review