Stitch Screen Generator

The skill explicitly routes project, screen, and asset operations through Stitch MCP and declares tool access, but it also grants Bash, Write, and web_fetch, requires downloading remote HTML/screenshots, and can overwrite local files. User confirmation, least-privilege rationale, data-flow disclosure, sensitive-data handling, rollback, and source verification are incomplete, so points are deducted. No malware, credential theft, covert exfiltration, or destructive default red line is evident.

The document gives fairly detailed steps, parameter examples, and iteration guidance for four main flows, but lacks diagnosable handling for missing projects, tool failures, invalid outputs, download errors, and metadata conflicts. The use of `run_command` is also inconsistent with the declared tool names. Static evidence cannot exceed the 10-point reliability ceiling, so points are deducted.

The scenarios are clear: text generation, image generation, editing, and variants, with basic inputs and device types. However, non-fit boundaries, exclusion triggers, Chinese-language support, and network reachability are not specified. The core function depends on Stitch MCP, credentials, and potentially overseas services, with no mainland-China availability guidance, so points are deducted.

The skill has layered documentation with workflows, prompt templates, references, and an example; repository context supplies Apache-2.0, a security policy, and limited CI information. The selected skill lacks versioning, changelog, named maintenance responsibility, update path, troubleshooting guidance, and concrete dependency installation notes, while cross-skill dependencies remain only partially operationalized, so points are deducted.

The goal, workflow, expected Stitch output components, and local asset destinations are mostly clear, giving the happy path direct value. However, downloads, metadata updates, and result verification depend on unspecified tool behavior and external services, and no representative outputs are verifiably demonstrated. Under static calibration the score remains at or below 7, so points are deducted.

The selected file and its referenced materials are auditable, and the repository contains a real CI workflow. That workflow validates react-components rather than generate-design, with no target-specific tests, third-party execution evidence, or corroborating sources. Only limited verifiability credit is justified.