The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses and reads raw conversation transcripts from ~/.claude/projects/. These transcripts contain the full history of user interactions with the AI, which may include sensitive data, code, or private information. While the data is used locally to calculate costs, reading these paths represents exposure of sensitive user history.
[PROMPT_INJECTION]: The skill is subject to indirect prompt injection risks because it processes raw conversation data which is then summarized for an AI subagent to perform analysis.
Ingestion points: Raw .jsonl transcript files are ingested from ~/.claude/projects/ via the analyze-usage.js script.
Boundary markers: The agent-prompt-template.txt does not specify the use of clear delimiters or instructions for the subagent to disregard instructions embedded within the processed chat history.
Capability inventory: The background agent has access to Bash, WebFetch, Read, and Edit tools, allowing for command execution and file system modification.
Sanitization: Although a --private flag is available to remove prompt text from the final HTML report, the analysis subagent still processes the raw, unvalidated transcript data.
[EXTERNAL_DOWNLOADS]: When encountering an unknown model, the skill is instructed to fetch pricing data from platform.claude.com. This information is used to dynamically update the local model-pricing.json configuration file via the Edit tool.
[COMMAND_EXECUTION]: The skill executes multiple shell commands to manage its analysis pipeline, including running Node.js scripts to prepare data, build reports, and launch the dashboard in the default system browser.

usage-view