The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill reads and processes user-provided journal entries from a local file, creating a surface for potential indirect instructions. The risk is minimized by the skill's limited capabilities and structured processing.
Ingestion points: Content is read from the journal file (defaulting to ~/research/journal.md) during the Monthly Review and via the read command in scripts/journal.py.
Boundary markers: The script enforces a specific header structure using regex (## YYYY-MM-DD kind: title) to delimit entries.
Capability inventory: The skill is restricted to reading and appending to a single local file. It does not have network access, external tool execution (beyond the bundled script), or system modification capabilities.
Sanitization: The logic validates the structure of entry headers but processes the textual content of journal fields as raw strings.
[COMMAND_EXECUTION]: The skill uses bundled Python scripts for its primary functionality and testing.
The agent is instructed to use python3 scripts/journal.py for journaling tasks.
The file scripts/test_journal.py uses subprocess.run() to execute the journal script for regression testing. This execution is limited to the bundled code using the local Python interpreter and does not incorporate untrusted input into shell commands.

credo-journal