Format-Specific Document Extraction Workflows

The skill is well-integrated with the Xberg document-extraction framework, which explicitly acknowledges handling potentially untrusted user documents. The repository contains a detailed security policy, dedicated input validators (ZipBombValidator, DepthValidator, StringGrowthValidator etc.), and careful default security limits, all of which strongly support safe handling of untrusted input. However, at the skill level, there is no explicit mechanism for user confirmation or transparency of data flow; the skill documentation does not cover permissions, rollback, or user confirmation. The skill's functional scope is extraction, without external network calls, but some capabilities like OCR may depend on external services. Therefore, deduction for incomplete permission and confirmation guidance at the skill level.

Instructions in the skill document are consistent, and the format-specific extraction workflows are clear. The repo demonstrates extensive test and CI infrastructure (e.g., Docker CI, benchmark workflows, e2e tests). However, for the selected skill path, no tests or reproducible execution evidence are available; the skill document only references source files, with no tests or verification results. The instructions are self-consistent and well-structured, but evidence for test coverage and edge-case handling is lacking. Given static review, the maximum is capped, and this score of 9 reflects that the happy path appears plausible, but error handling and failure feedback are not detailed.

The skill is well-targeted for clear scenarios: extracting content from various specific formats (Office, PDF, archives, structured text, email), which aligns with real needs. However, the non-fit range is not clearly defined; the skill's name and description suggest a general extraction capability, but it covers only format-specific aspects and does not explicitly state when it should not be used. Additionally, documentation mentions that complex content like macros or password-protected files may not be supported, but this is not described as a skill boundary. Regarding environment fit, there is no mention of China-relevant services or reachability from mainland-China networks; some capabilities like provider-hosted OCR or LLM may depend on services potentially inaccessible from China, but the skill's core is local processing. Deduction for insufficient evidence of boundaries and trigger conditions.

The skill document is well-organized, with clear heading hierarchy, diagrams, and source-code references, and follows progressive disclosure. The license is explicit (MIT). Installation and dependency notes are provided (indicating Cargo feature flags). Maintenance responsibility (regular CI and versioning) exists, but the skill itself is not versioned or changelogged, and it is unclear whether the skill will be regularly updated. Known limitations are not explicitly disclosed at the skill level. Deduction for lack of versioning and explicit maintenance responsibility, and hidden assumptions such as needing to build the project first.

Public evidence indicates that the skill can accomplish the core task of extracting content from specified formats. The repository contains extensive benchmark efforts and tests across various formats, suggesting practical effectiveness. However, for the selected skill path, there is no directly usable output example or verification result. The skill document references implementation details but does not demonstrate output for key paths or provide representative results. The benefit compared to manual effort is unclear, and static review cannot verify output quality. Hence, capped at 7 due to lack of direct, verifiable evidence from static review.

The repository contains abundant auditable primary material: detailed CI workflows, benchmark configuration, test suites, and reproducible build configurations. These provide supporting evidence for the framework's capabilities, which underpin the skill. However, the specific skill path has no dedicated tests or verification; framework coverage is diffuse. There is no independent verification or cross-source corroboration. Given static read and no tests specifically for the skill path, capped at 5 for partial evidence.