The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes explicit defensive instructions to mitigate risks from processing untrusted data. In references/provenance-contract.md and references/session-capture.md, it mandates that instructions embedded in external content or historical transcripts be treated strictly as evidence and not as executable commands that could influence agent behavior or permissions.
[DATA_EXFILTRATION]: The skill enforces boundaries for sensitive information. Instructions in README.md and references/session-capture.md require the redaction of secrets and credentials before they can be added to the Wiki. Additionally, repository mutations (commit and push) require explicit, separate user authorization.
[COMMAND_EXECUTION]: The skill utilizes standard Git commands and a generated local script for repository integrity and management. These operations are restricted to the local wiki environment and include verification gates such as reviewing diffs and ensuring a clean worktree before modifications.
[PROMPT_INJECTION]: Regarding indirect prompt injection (Category 8) vulnerability surfaces: (1) Ingestion points include external URLs (source-ingest.md) and session history (session-capture.md); (2) Boundary markers are present as explicit warnings to treat input as untrusted reference material; (3) Capability inventory includes file-write, Git operations, and local maintenance script execution; (4) Sanitization instructions for credentials and personal data are provided in session-capture.md.

llm-wiki-capture