Frameloop has 20 tools with basic schemas and descriptions, but significant quality gaps limit production readiness. All tools have descriptions (10-200 chars, mostly adequate), and input schemas are present with type definitions. However, parameter descriptions are sparse or missing context; output schemas are undocumented; error handling is absent; and tool composition shows weak chaining (e.g., render_still outputs to disk but no tool reads rendered frames). No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels. Naming is verb-forward (render_*, get_*, extract_*) but some tools conflate concerns (render_video + render_gif should be one parameterized tool). Security: no visible input validation, path traversal risk in file operations, and no audit logging.

• Document output schemas for all tools. Specify field names, types, and structure. E.g., render_still returns {image_path: string, width: int, height: int, frame_number: int}.
• Add error handling with recovery guidance. E.g., 'Composition not found. Call get_compositions() to list available compositions.' instead of raw exceptions.
• Validate and sanitize all file paths. Reject paths with '..' or absolute paths outside project directory. Use pathlib.Path.resolve() and assert it's within allowed root.
• Merge render_video and render_gif into a single render_output tool with format enum (video|gif). Reduces cognitive load and prevents LLM confusion.
• Add tool annotations: mark render_* and scaffold_project with destructiveHint=true; mark get_* and list_* with readOnlyHint=true. Enable LLM reasoning about side effects.