The Agent Skills Directory

[PROMPT_INJECTION]: The code snippets in SKILL.md demonstrate patterns for indirect prompt injection.
Ingestion points: Untrusted data is brought into the agent's context through web scraping in Step 6 (fetch_webpage), message monitoring in Step 9 (DiscordChannel), and memory retrieval in Step 17 (MemoryStore.recall).
Boundary markers: The tutorial's prompt templates, such as those in Step 5 and Step 17, do not include delimiters or instructions to ignore instructions embedded within the ingested data.
Capability inventory: The agents described have tool-calling capabilities and network access, providing a path for injected instructions to affect the system.
Sanitization: The code lacks mechanisms for validating or sanitizing external content before it is processed by the language model.
[EXTERNAL_DOWNLOADS]: SKILL.md contains instructions for downloading external code and packages.
GitHub Repository: The instructions include cloning a repository from https://github.com/czl9707/build-your-own-openclaw.git for the complete tutorial code.
Package Installation: The guide directs users to install dependencies from requirements.txt using the pip package manager.

build-your-own-openclaw-agent-tutorial