The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to extract 'new decisions', 'corrected rules', and 'new methods' from conversation history and code diffs to update local documentation (e.g., CLAUDE.md, HARD_RULES.md). This creates a feedback loop where instructions provided during a chat session can become persistent rules for the agent in future sessions.
Ingestion points: Dialogue text, git diff output, and git log metadata.
Boundary markers: The skill requires a 'Proposal -> Approval -> Write' workflow, providing a human-in-the-loop checkpoint, though it lacks automated sanitization for injected instructions.
Capability inventory: The skill possesses file-write capabilities across the repository, including project configuration and agent-instruction files.
Sanitization: No explicit sanitization or filtering of the ingested dialogue is mentioned beyond the manual approval process.
[COMMAND_EXECUTION]: The skill uses shell commands to gather context about the project state.
Evidence: The instructions explicitly call for running git diff and git log on the repository to identify changes for the 'Collection' phase.
[DYNAMIC_EXECUTION]: The skill delegates complex processing tasks to 'child agents' (子代理).
Evidence: The workflow involves spawning a separate agent context to perform consistency checks and file writes, which uses a 'burnable' window to manage context length. While this is a common architectural pattern, it involves dynamic task delegation.

cyxj-remotion-evolve