The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill directs the agent to install the garmin-pp-cli and garmin-pp-mcp tools from the vendor's official NPM package (@mvanhorn/printing-press-library) and GitHub repository (github.com/mvanhorn/printing-press-library).
[DATA_EXFILTRATION]: Includes a --deliver webhook:<url> feature that enables the agent to POST command results, which may contain sensitive health information, to arbitrary remote endpoints. It also includes a feedback mechanism that can be configured to send local logs to a remote endpoint.
[COMMAND_EXECUTION]: The skill is designed around driving the garmin-pp-cli binary. It provides recipes for running complex subcommands, including a sql command that executes direct queries against a local SQLite database.
[DYNAMIC_EXECUTION]: Implements a 'playbook' system that dynamically assembles and replays command sequences ('choreography') based on local journals of past sessions. This allows the agent to execute multi-step procedures stored in its local data store.
[INDIRECT_PROMPT_INJECTION]: The 'learning loop' (recall and teach functions) creates a vulnerability surface by having the agent retrieve and follow instructions or playbooks from a local data store, which could be influenced by external data if the store is not properly managed.
Ingestion points: The recall command ingests instructions from a local journal and the data.db SQLite archive.
Boundary markers: The instructions suggest using file redirection (QUERY=$(cat ...)) to prevent shell injection, but do not define delimiters to prevent the agent from being misled by the content of the recalled instructions.
Capability inventory: The skill has the capability to read files, execute shell commands, and perform network POST requests.
Sanitization: The skill advises the agent to strip personally identifiable information (PII) before recording new 'learnings,' but does not specify validation or sanitization for data retrieved during the recall process.

pp-garmin