The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill identifies improvement signals by scanning the current conversation history. If an external data source (like a webpage or document read by the agent) contains malicious instructions designed to look like user corrections or preferences, the skill could be tricked into proposing unsafe code changes or commands.
Ingestion points: Conversation history, including previous assistant outputs and user inputs (SKILL.md step 2).
Boundary markers: Absent; the skill does not explicitly use delimiters to separate instructions from the data being analyzed for signals.
Capability inventory: File writing (SKILL.md step 4), shell command execution for tests and Git operations (SKILL.md step 4), and network operations via Git push (SKILL.md step 4).
Sanitization: Relies on human-in-the-loop verification via multiple approval prompts before changes are applied, committed, or pushed.
[COMMAND_EXECUTION]: The skill invokes shell commands to manage version control and validate changes.
Evidence: Step 4.1 uses read-only commands to resolve Git root, remote, and branch. Step 4.4 and 4.5 execute 'focused tests', 'documented skill lint/validation', and 'git commit/push'.
[DYNAMIC_EXECUTION]: The skill creates or modifies executable scripts and regression tests, then immediately executes them to verify the fix.
Evidence: Step 4.4 describes applying changes to bundled scripts and hooks, then running the updated tests to verify the new behavior.

reflect