The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from scraped competitor web pages and user-pasted positioning canvases.
Ingestion points: External data is ingested through scripts/connectors/firecrawl.py (scraping) and scripts/connectors/tavily.py (searching), as well as user-supplied text.
Boundary markers: The instructions explicitly command the agent to treat these sources as untrusted input and to ignore any embedded instructions (referencing SECURITY.md).
Capability inventory: The skill has the ability to perform network searches, read local memory files, and write to event logs via registry-events.py.
Sanitization: The skill relies on natural language instructions to the agent to prevent execution of embedded commands within the ingested data.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill initiates network requests to external services for competitive research.
Evidence: It utilizes scripts/connectors/tavily.py and scripts/connectors/firecrawl.py to fetch information from external domains. These are well-known search and scraping services used for standard data retrieval and do not involve the exfiltration of sensitive local credentials or system files.
[COMMAND_EXECUTION]: The skill interacts with the local file system and internal event buses using specific scripts.
Evidence: It executes registry-events.py with an operation: propose argument to update event logs (memory/events/*.ndjson). These operations are scoped to the skill's marketing workflow and do not include arbitrary shell command execution or privilege escalation.

positioning-truth-tracer