Loading ToolBench
AI-native cultural art creation & evaluation
23 tools with mixed quality. Naming is generally verb-forward (create_, generate_, evaluate_, list_, get_, search_, layers_*) and mostly clear. Descriptions present but often terse (many 50-150 chars, below the 194-char baseline). Schemas visible in audit scripts but incomplete: most tools lack explicit type definitions for all parameters, and output schemas are not documented. Error handling and recovery guidance absent. No evidence of parameter validation rules, enum constraints, or actionable error messages. Tool composition is reasonable (layers_* form a coherent subsystem), but many tools lack the depth of documentation expected for production LLM agents.
Archive the current session.
Parse a natural language brief into structured intent components.
Compose a prompt from a design document.
Generate and evaluate an image in one call — convenience wrapper over generate_image + evaluate_artwork. Returns image + L1-L5 scores in a single round-trip. For fine-grained control (retry loop, custom prompts, separate scoring), use generate_image and evaluate_artwork separately.
Score an image on L1-L5 cultural dimensions — returns scores, rationales, recommendations. Use after generate_image or layers_composite to decide: accept, retry, or edit layers. Combine with view_image to correlate visual inspection with numeric scores.
Generate multiple concept variations from a prompt.
Output schemas not documented. Tools like evaluate_artwork, layers_composite, and generate_image return complex results (scores, rationales, images, metadata) but no schema is visible. LLMs cannot plan downstream calls or extract fields reliably.
Parameter type definitions incomplete. Many tools (e.g., layers_transform, layers_edit, sync_data) show parameter names and descriptions but lack explicit JSON Schema type declarations (string, number, boolean, array, object).
Descriptions too terse for LLM selection. 11 tools have descriptions under 60 characters (list_traditions, view_image, layers_list, unload_models, sync_data). Scores capped at 50.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Generate an image from a prompt using specified provider.
Get detailed guide for a specific cultural tradition.
Inpaint a region of an artwork with specified instruction.
Composite layers back into a single image.
Edit a layer (toggle visibility, modify properties).
Evaluate all layers in an artwork.
Export layers to various formats.
List all layers in an artwork.
Paste a layer back onto the source image.
Redraw a layer with new instruction.
Split an artwork into semantic layers.
Transform a layer (translate, rotate, scale).
List all available cultural traditions with metadata.
Search traditions by keywords or attributes.
Synchronize data with remote storage.
Unload models from memory to free resources.
View or retrieve an image file.
No enum constraints on provider/tradition parameters. Tools accept 'provider' (mock|gemini|nb2|openai|comfyui) and 'tradition' (cultural identifiers) as free-form strings. No enums visible in schemas. LLMs will hallucinate invalid values.
No error handling or recovery guidance. Tools lack descriptions of failure modes, retryability, or next steps. E.g., if generate_image fails due to invalid provider, the LLM has no guidance on what to try next.