Conversation Archivist

The skill declares only Read, Write, Edit, Glob, and Grep, while its procedure requires bash/python execution, CLI transport, or MCP writes; this permission/data-flow mismatch costs points. It includes destination checks, locking, filename sanitization, and overwrite confirmation, showing partial scoping and recovery awareness. However, it does not explicitly confirm the save action itself, address sensitive conversation handling, disclose external transport in the user flow, or provide transactional rollback; index, log, and hot-cache updates can be partial.

The routing, transport selection, locking, note-type table, and naming rules are fairly concrete, and repository CI/tests provide limited static evidence for shared infrastructure. There are no tests for the save workflow itself, including routing conflicts, lock-release failure, partial writes, unavailable CLI/MCP, duplicate pages, or abnormal conversations. The declared tool set also conflicts with required commands, so the static score remains conservative.

Trigger phrases, note types, folders, project-vault versus personal-vault precedence, and four methodology modes are described. Non-fit boundaries, execution of the skip decision, environment prerequisites, and end-to-end Chinese-language behavior are under-specified. The local Markdown/Obsidian target is broadly compatible with Chinese users, and the skill is not shown to depend entirely on unreachable overseas services.

The skill is well layered across transport, modes, concurrency, decision rules, workflow, templates, writing style, and skip criteria; repository context supplies MIT licensing, security policy, CI, and version signals. The skill itself lacks explicit maintenance ownership, changelog, installation/dependency notes, troubleshooting, and stable parameter governance. References to other skills and global CLAUDE.md are not evidence contained in the selected path, so points are deducted.

The core outcome is clear: convert the current conversation into a structured, searchable Markdown note with frontmatter, links, index, log, and hot-cache updates. The extraction and linking rules support the main happy path and the resulting artifact is intended to be directly usable. However, success depends on unprovided vault state, CLAUDE.md, scripts, and transport integrations; there is no save-specific output sample or verification, and multi-file updates may be incomplete, so the score stays within the static ceiling.

README, LICENSE, SECURITY, CI, and several general tests provide auditable repository evidence. CI validates some shared scripts and frontmatter, but there is no dedicated save/SKILL.md test, revision-level behavior evidence, representative saved-note fixture, or independent corroboration. Marketing claims about broader capabilities do not verify this selected skill path, warranting only limited credit.