The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill interacts with external project data through olcli pull and olcli download commands (ingestion points identified in SKILL.md). Since Overleaf projects are often shared, the LaTeX source code represents untrusted data. There are no boundary markers or instructions to sanitize or ignore embedded prompts within these files. The agent has significant capabilities, including network access and local file writing through olcli push and olcli sync, creating a surface for indirect prompt injection.
[EXTERNAL_DOWNLOADS]: The skill installs the olcli utility from the author's infrastructure, specifically using the aloth/tap Homebrew tap and the @aloth/olcli npm package as seen in README.md and scripts/install.sh.
[COMMAND_EXECUTION]: The skill relies on the execution of a local shell script (scripts/install.sh) and the olcli binary to perform project management, compilation, and synchronization tasks.
[CREDENTIALS_UNSAFE]: The authentication flow uses the overleaf_session2 browser cookie. While documented as the intended mechanism for olcli, this involves the management of sensitive session credentials which are stored in local files such as ~/.config/olcli-nodejs/config.json or .olauth as described in references/API.md.

overleaf