The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to fetch data from external, untrusted web sources (e.g., DefiLlama, CoinGecko, Glassnode, and general web results for market trends) to populate DeFi, on-chain, and IPO analysis reports. This creates a surface where malicious content on these platforms could influence the agent's behavior.
Ingestion points: WebSearch results for DeFi protocol metrics, on-chain analytics, market trend discovery, and IPO prospectus fallsbacks (references/defi-yield.md, references/onchain.md, references/finance-calendar.md, references/hkipo-analysis.md).
Boundary markers: The instructions lack specific delimiters or "ignore instructions" warnings to prevent the agent from following directives potentially embedded in the retrieved web content.
Capability inventory: The skill possesses significant capabilities, including the execution of the longbridge CLI to fetch account-level data (portfolios, positions) and market data across numerous reference files.
Sanitization: There are no explicit instructions for the agent to sanitize, validate, or filter the content retrieved from external URLs before integrating it into its reasoning or output.
[METADATA_POISONING]: The skill's YAML frontmatter in SKILL.md identifies the author as "longbridge," whereas the provided skill context identifies the author as "nickbuddy2046." While this is consistent with the skill's focus on the Longbridge CLI, the discrepancy in identity attribution represents a metadata inconsistency.
[METADATA_POISONING]: Several files (e.g., references/hkipo-analysis.md, references/finance-calendar.md) reference internal documentation files such as references/scoring.md, references/output-template.md, and references/data-fetching.md which are not included in the skill package. This results in incomplete instructions for the agent's logic.

stock-research