The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes improvement entries from .turbo/improvements.md, which acts as an external data source.
Ingestion points: The file .turbo/improvements.md is parsed in SKILL.md Step 1 to extract summaries, types, and categories.
Boundary markers: Data is structured using Markdown headers (###) and specific metadata fields (Type, Category, Where, Why).
Capability inventory: The skill uses shell commands for repository path resolution and delegates task execution to other specialized skills ($implement, $investigate, $turboplan).
Sanitization: While there is no explicit string escaping, the skill mandates a 'Validate and Classify' phase (Step 2) where the agent must confirm the existence of issues by reading the actual source code, significantly reducing the risk of accidental obedience to malicious instructions in the backlog.
[COMMAND_EXECUTION]: The skill utilizes git rev-parse to accurately resolve the repository root and detect linked worktrees. These are standard, benign operations for a developer tool intended to work within a Git repository.

implement-improvements