The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill inherently processes untrusted human input from Google Docs tabs and comments as part of its document refinement cycle.
Ingestion points: Document content and comments are read using gws-axi docs read and gws-axi docs comments as described in SKILL.md and references/google-docs.md.
Boundary markers: The workflow uses a structured 'grid' table to strictly separate human-controlled status rows from agent-controlled document content.
Capability inventory: The agent possesses the capability to write/edit Google Docs tabs and perform local git operations (hash-object, commit).
Sanitization: The skill instructions mandate a robust 'body-to-body' diffing strategy against the original git source (using git cat-file) to identify human modifications. The agent is explicitly instructed to treat these changes as content signal rather than behavioral instructions.
[COMMAND_EXECUTION]: The skill utilizes local system commands to automate document lifecycle management and version control.
Evidence: The scripts/render-draft.mjs utility executes git hash-object, git rev-parse, and git remote via Bun.spawnSync to generate provenance metadata.
Evidence: SKILL.md and references/google-docs.md provide procedures for the agent to use the gws-axi CLI for interacting with Google Docs and the git CLI for repository maintenance.

doc-drafting