Loading ToolBench
The PowerPoint Creator server has 6 tools with mixed quality. Tool naming starts with action verbs (create_, add_, generate_) which is correct. However, descriptions vary significantly in completeness and specificity. All 6 tools have input schemas with type definitions and parameter descriptions, which is good. The major issues are: (1) descriptions lack guidance on WHEN to use each tool vs. similar alternatives, all 6 add_*_slide tools do similar things but lack disambiguation; (2) parameter defaults (title=None, subtitle=None) could cause silent failures or unexpected output; (3) no structured output schemas are documented, tools return simple strings rather than structured objects with fields the LLM can reliably parse; (4) error handling is basic (returns error strings) without recovery guidance; (5) no tool annotations (destructiveHint, idempotentHint, readOnlyHint) to help LLMs reason about side effects. The generate_svg tool is notably underdocumented for its complexity (requires rsvg-convert external dependency, file I/O, subprocess calls). The create_presentation tool is simple and clear. Overall, these are functional tools but lack polish for production LLM agents.
Add a content slide with bullet points.
Add a slide with an image and title.
Add a section divider slide with a large title.
Add a title slide to the presentation.
Create a new empty PowerPoint presentation with a title.
Takes final <svg>…</svg> text, writes to /tmp/, rasterizes via rsvg-convert, and inserts as a picture into a new slide.
Multiple tools (add_title_slide, add_content_slide, add_section_slide, add_image_slide) perform similar operations (add a slide to a presentation). Descriptions do not disambiguate when to call which. An LLM cannot tell whether to use add_content_slide vs add_title_slide from descriptions alone.
generate_svg description does not document the external dependency (rsvg-convert), does not explain what happens if rsvg-convert is not installed, and does not provide recovery guidance. Error handling returns 'Error: rsvg-convert not found. Please install librsvg...' which is actionable but not integrated into the description.
Tools return plain string responses (e.g., 'Added title slide at position 3', 'Error: Presentation not found'). These are unstructured and force LLMs to parse text to extract the slide position, success/failure status, or presentation ID. Output should be structured JSON with typed fields (position: int, success: bool, prs_id: string).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 50 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
No tool annotations present. Tools that modify presentations (all 6 tools) should declare destructiveHint=true so LLMs know they have side effects and require careful planning. generate_svg performs file I/O and external subprocess calls, these should be marked to guide LLM reasoning.
add_title_slide and add_content_slide have optional parameters with default=None (subtitle, background_color). When omitted, these parameters silently produce minimal output. Descriptions do not explain the consequences of omitting them or what the default behavior is.
Error responses are inconsistent. Some return 'Error: Presentation not found', others return 'SVG→PNG error: ...'. No categorization (retryable vs. user-fixable vs. fatal) is provided. LLMs cannot determine whether to retry, ask the user, or abort.
The create_presentation tool returns a presentation ID (e.g., 'created_ppt_abc123'), but subsequent tools require prs_id as a string parameter. The response does not explicitly state that this ID must be passed to subsequent operations, forcing the LLM to infer the connection.