The Agent Skills Directory

[CREDENTIALS_UNSAFE]: A hardcoded authentication token identified as 'Master base token' (Yt2WbfTeQa0tjQsjMUwlfaEvgXb) is present in references/changing-the-base.md. While the skill uses this for identity comparison to prevent unauthorized writes to a production database, hardcoding master credentials within the skill instructions exposes sensitive access keys to any environment where the skill is processed.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from Meta Ads performance bases, creating an attack surface where malicious data could influence agent behavior.
Ingestion points: The skill ingests performance data, lead records, and account metadata from external Lark/Bitable tables (SKILL.md Step 0, references/opening-checks.md).
Boundary markers: While the skill implements logic-based 'opening checks' to validate data freshness and consistency (SKILL.md), it does not utilize prompt-level delimiters or 'ignore instructions' warnings to prevent external data from being interpreted as malicious instructions.
Capability inventory: The agent possesses the capability to write verdict snapshots to the local filesystem vault (references/verdict-snapshot.md) and execute database operations via lark-cli (SKILL.md, references/changing-the-base.md).
Sanitization: The skill lacks explicit sanitization or escaping mechanisms for data retrieved from the external database before it is interpolated into the agent's context for analysis.

breakthrough-performance-analysis