The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes conversation history to generate a handoff document for subsequent agent sessions. While this creates a potential surface for instructions to persist across sessions, the skill includes specific mitigations to handle untrusted content.
Ingestion points: Reads the current conversation context to summarize work progress.
Boundary markers: Uses Mermaid diagrams and 'Beliefs downgraded' sections to structure the output and differentiate between verified facts and assumptions.
Capability inventory: Includes file writing, git operations (status, commit, push), and claude CLI commands.
Sanitization: Explicitly mandates the redaction of API keys, tokens, passwords, and personal data before output generation.
[COMMAND_EXECUTION]: The skill leverages shell commands for repository management and starting new cloud sessions.
Evidence: Uses git status, git commit, git push, and claude --cloud/claude --teleport commands.
Control Flow: The skill is configured with disable-model-invocation: true and instructions emphasize that commands must only be run after explicit user approval ('assisted git, never automatic').
[DATA_EXFILTRATION]: The skill facilitates pushing documents to a remote GitHub repository when using the cloud destination.
Analysis: This is the intended primary purpose of the skill to allow cloud-based agents to access the handoff context. The risk is mitigated by the requirement for manual user confirmation and the secret redaction policy.

handoff