The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is configured to ingest arbitrary content from branch diffs, staged/unstaged changes, and untracked files to generate demo stories. This creates a surface for potential indirect prompt injection from malicious content within the analyzed source code. The skill addresses this by providing explicit instructions for the agent to filter and redact sensitive information.
Ingestion points: SKILL.md (Contract point 1) and REFERENCE.md (Ground the story point 1) instruct the agent to inspect the whole diff from the merge-base, including all file states.
Boundary markers: The instructions include clear directives to "exclude secrets/private/customer PII" and "Redact secrets, tokens, private URLs, customer PII, internal-only data".
Capability inventory: The skill possesses the ability to execute shell commands (bunx, open), write to the local filesystem (demos/<slug>/), and perform Git operations (commit, push, PR).
Sanitization: The skill relies on the agent's instruction-following capabilities to sanitize demo content and PR descriptions before publication.
[EXTERNAL_DOWNLOADS]: The skill uses bunx to fetch and run create-video@latest and the Remotion render CLI. These are standard, well-known developer tools provided by the Remotion ecosystem.
[COMMAND_EXECUTION]: The skill executes local shell commands for rendering video files (bunx remotion render) and for revealing files in the OS file explorer (open -R). These commands are standard for developer productivity workflows and are used within the context of the current project repository.

demo