1How It Works & Mechanism
A pure prompt-orchestrated "knowledge base engineering" skill: leverages a fixed set of files (SOURCE_OF_TRUTH.md / AGENTS.md / CLAUDE.md / CODEBUDDY.md / local README.md files) to structure regular folders into navigable paths for agents in every turn, without introducing vector databases, RAG, or external databases.
skills/dbs-knowledge/SKILL.md用户只需要理解「知识库」三个字。`SOURCE_OF_TRUTH.md`、`AGENTS.md`、`CLAUDE.md`、`CODEBUDDY.md` 和宿主兼容规则都由你在后台处理。
The "Default Prohibitions" explicitly state: never require users to install databases, vector stores, embedding services, or RAG systems. It relies entirely on file conventions and prompt discipline rather than indexing technologies.
The core artifact is a traceable invocation chain: Entry Rules → SOT Navigation → Directory README → Raw Files; navigation only provides directions without storing facts, and answering questions strictly requires reading the original raw files.
skills/dbs-knowledge/SKILL.mdSOURCE_OF_TRUTH.md：去哪里找、放在哪里、多个版本以哪个为准
There is a strict constraint in the document: "Never answer questions based solely on summaries in the navigation."
SKILL.md routes intents into four core internal modes (Create, Store, Use, Inspect); when handling multiple intents, it processes them in dependency order (e.g., ingest files first, then answer questions dependent on those files).
skills/dbs-knowledge/SKILL.md用户同时提出多个意图时，按依赖顺序处理。例如先收录新文件，再回答依赖该文件的问题。
The routing table maps "SOT too large / navigation too long" to the "Inspect and Govern" mode.
Write operations adopt a tiered confirmation threshold: read-only audits execute freely; minor modifications with explicitly identified targets require only a single-sentence confirmation; operations involving move, copy, rename, overwrite, delete, or batch actions must provide a preview before execution.
skills/dbs-knowledge/SKILL.md用户确认后再写入。查询知识库和只读健康检查不需要确认。
This serves as the primary safety mechanism to preserve file integrity while allowing modifications.
Safety boundaries are hardcoded in the "Default Prohibitions" list: no uploading to third parties, no modifying original text, no guessing authoritativeness, no unconfirmed batch operations, and no reading secrets, browser data, or chat databases.
skills/dbs-knowledge/SKILL.md不上传用户文件到第三方平台。
The same list includes: "Never read secrets, passwords, browser data, chat databases, or other sensitive content" and "Never move, rename, overwrite, or delete files in batch without explicit confirmation."
Lightweight updates use an independent fast path: read only the SOT, update necessary entries, and strictly forbid full-vault scans, modifying Agent entry points, or reorganizing unrelated files in passing.
skills/dbs-knowledge/SKILL.md不扫描整个知识库；
Comes with a fixed receipt format ([Modified] / [Added] / [Deleted] change logs) that reports only actual changes made in the current session.
The governance (slimming) mode defines four rules—Keep, Merge, Submerge, and Delete—and mandates backing up the original navigation to .trash/ inside the knowledge base before writing, ensuring backups never become the authoritative version.
skills/dbs-knowledge/SKILL.md先把原导航复制到知识库内的安全备份位置，推荐 `.trash/YYYY-MM-DD_SOURCE_OF_TRUTH_优化前.md`
Followed immediately by: "Backups will never become the active authoritative version" to prevent backups from being mistaken for the latest source.
Mandatory post-governance verification strictly distinguishes "path existence" from "content authority": dynamic facts (amounts, prices, statuses, assignees) must be cross-checked against raw files.
skills/dbs-knowledge/SKILL.md路径验证通过不等于内容权威。
Prevents false-positive sign-offs where an optimization is deemed successful simply because links are not broken.
2Core Capabilities
01Knowledge Base Creation: Supports empty directories and existing documents, generating SOURCE_OF_TRUTH.md and a minimal folder hierarchy (including a low-friction 00-Inbox/ folder).
02Ingestion: Assesses type, source, date, and version, suggests target paths, defaults to copying outside files to preserve originals, and avoids overwriting duplicate names.
03Invocation: Navigates to a minimal set of relevant files, reads raw content to answer queries, and outputs "Source Files + Freshness and Information Gaps".
04Health Checks: Audits path validity, multiple "latest version" conflicts, overlapping directory scopes, 00-Inbox/ backlogs, and dangling references, triaged into Immediate, Recommended, and Keep As-Is tiers.
05SOT Tiered Governance (Slimming): Audits topic density, applies Keep/Merge/Submerge/Delete rules, and offloads granular details to local directory READMEs.
06Multi-Client Entry Adaptation: AGENTS.md holds cross-platform rules, CLAUDE.md serves as a thin entry point via @AGENTS.md, CODEBUDDY.md is optional, and Grok bridge requires user_invocable: true.
07First-Use Onboarding Generator: Dynamically generates 4 directly copyable example prompts matching the actual current directory structure, strictly forbidding generic placeholder templates.
08Re-bloat Prevention Mechanism: Updates the top-level navigation only when new materials change authoritative sources, directory responsibilities, version relationships, or high-frequency entry points.
6Verdict & Scenarios
Breaks down the transition from "raw folder to agent-ready knowledge base" into four actionable modes (Create/Store/Use/Inspect) with explicit completion criteria, offering engineering rigor far beyond typical memory or knowledge prompts.The governance mode addresses the common degradation issue of knowledge base bloat through Keep/Merge/Submerge/Delete rules, anti-bloat provisions, and post-governance verification.Systematic safety design: unrestrained read-only audits, tiered write confirmations, pre-execution previews with .trash backups for batch actions, exclusion of sensitive directories by default, and fail-fast behaviors that explain configuration issues rather than degrading silently.User-friendly transparent terminology: uses accessible terms ("Knowledge Base", "Navigation", "Current Active Version", "To Organize") and generates 4 tailored example prompts based on real local directories upon first use.Clear boundary discipline: specifies distinct responsibilities across 4 sibling skills in explicit sections, handing off out-of-scope requests back to /dbs instead of stringing together bloated agent chains.Best for: Ideal for individual knowledge workers and small teams managing scattered local documents (notes, contracts, course drafts, client dossiers, legacy versions) who want agents to reliably locate files and cite sources in every turn; particularly well-suited for setups avoiding vector DBs and RAG in favor of plain files and structured conventions. Also serves as a solid foundation for other skills (establishing navigation before content-focused skills operate on the repository).
Not for: Not suitable for scenarios requiring technical retrieval guarantees (embeddings or full-text search indexing)—its retrieval ceiling equals the model's ability to navigate files; not recommended for massive data scales (tens of thousands of files) or multi-user concurrent writes; not intended for purely read-only needs (usable, but its primary value lies in creation and governance); not suitable for unattended automation, as all write operations require user confirmation.