Expo SDK Upgrade Assistant

The material discloses the upgrade, dependency, native-directory, and cache-clearing workflow and tells the agent to check for ios/android first. However, it includes potentially destructive or costly commands such as `npx expo prebuild --clean` and `rm -rf node_modules .expo` without requiring confirmation, backups, or rollback, and it does not fully explain network and dependency-install data flows. The repository supplies MIT licensing and source links, but the publisher is unverified by FollowSkills.

The procedure, reference files, and openai.yaml target are broadly consistent and cover CNG, bare workflows, deprecated packages, and migration caveats. There is no committed test suite for the key upgrade paths, no pinned environment assumptions, and no consistent failure-handling protocol; several checks are merely commands the user is told to run. Static evidence therefore supports only a score below the 10-point execution ceiling.

The audience and main scenarios are clear: SDK upgrades, dependency fixes, React 19, React Compiler, New Architecture, and deprecated-package migrations. Boundaries are still incomplete: project inputs, unsupported cases, and semantic trigger conditions are not fully defined, and the guidance relies on English Expo documentation and commands without Chinese support or mainland-China reachability information.

The SKILL.md, openai.yaml, and topic-specific references provide clear information architecture. Metadata includes name, version, and MIT license, while repository context shows installation guidance, version-check CI, and maintenance checks. The skill itself lacks a changelog, named maintenance responsibility, update path, explicit dependency/permission notes, comprehensive examples, and systematic troubleshooting; some versioned links and future-SDK claims may also age.

The skill provides copyable commands, migration tables, before/after code, and completion criteria, so it plausibly covers the core upgrade task. It contains no real project input/output, build result, or test evidence; users must still resolve project-specific conflicts, verify native behavior, and handle rollback. Under static calibration, effectiveness cannot exceed 7, so 7 is the conservative maximum.

Specific commands, code examples, reference links, and repository CI checks provide some auditability. The selected skill has no committed tests covering its key paths, no actual CI execution evidence, and no independent corroboration in the supplied material; external reference contents were not cross-verified here. A score of 4 reflects limited static traceability.