The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of external investment documents like earnings call transcripts and analyst reports into a personal wiki. This activity creates a potential surface for indirect prompt injection attacks. \n * Ingestion points: Files are read from the raw/ directory and compiled into wiki/ knowledge pages (referenced in SKILL.md and references/ingest_sop.md). \n * Boundary markers: The instructions mandate a "HITL 5 卡点" (Human-In-The-Loop) process, forcing the agent to stop and obtain user confirmation for takeaways, entity creation, and classification at five distinct stages. \n * Capability inventory: The skill possesses capabilities for local file reading/writing (via scripts/init_vault.py and scripts/lint-vault.py) and utilizes Git for version control and automation. \n * Sanitization: The skill relies on human oversight and manual verification instead of automated code-based sanitization for incoming text content. \n- [COMMAND_EXECUTION]: The skill utilizes local Python scripts (scripts/init_vault.py, scripts/lint-vault.py) and shell-based Git hooks (templates/pre-commit.snippet) to scaffold the project structure and validate markdown link integrity. These are standard utility scripts for maintaining a file-based knowledge system. \n- [EXTERNAL_DOWNLOADS]: The linter workflow uses the uv tool to load the pyyaml dependency for parsing frontmatter in the vault files. This process fetches the package from official registries, which is recognized as a well-known service and a standard development practice.

llm-wiki-setup