The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes local command-line tools to gather context for its explanations.
Evidence: Uses git diff, git ls-files, and git branch to analyze working tree changes.
Evidence: Executes a repository-local Python script located at .claude/scripts/code_graph to trace code dependencies and flow.
[PROMPT_INJECTION]: The instructions contain directives aimed at standardizing agent behavior across different platforms.
Evidence: Includes instructions to 'Ignore Claude-specific mode-switch instructions' and 'Never ask the user questions' to ensure a non-blocking, one-way explanation flow.
Evidence: Contains a 'Strict execution contract' demanding the agent follow the protocol exactly as written when invoked.
[DATA_EXPOSURE]: The skill accesses project metadata and source code to fulfill its primary purpose.
Evidence: Reads docs/project-config.json, plan.md, and various project reference files to synthesize explanations.
Evidence: Writes a persistent 'understanding ledger' to a git-ignored temporary directory (tmp/understand/) to track progress, avoiding modification of tracked source files.

understand