菜单
The Palace Protocol for interacting with MemPalace — wake-up sequence, query-before-assert pattern, memory stack usage, and when to use search vs KG vs navigation. Use when agents need guidance on how to interact with the palace correctly.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/KingInYellows/yellow-plugins --skill palace-protocol命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
| name | palace-protocol |
| description | The Palace Protocol for interacting with MemPalace — wake-up sequence, query-before-assert pattern, memory stack usage, and when to use search vs KG vs navigation. Use when agents need guidance on how to interact with the palace correctly. |
| user-invocable | false |
Defines how agents should interact with MemPalace to maintain data integrity and get the best results: wake-up sequence, core query-before-assert rules, memory stack layers, and tool selection guidance for search vs knowledge graph vs navigation.
Load when an agent needs to decide which mempalace tool fits the task, or at the start of a session that will draw from palace memory. Reference the "When to Use What" table to pick between search, KG, and navigation tools.
On first interaction with the palace in a session:
mempalace_status to load palace overview and protocol specQuery before asserting: Never state facts about a person, project, or
event without first querying the palace. Use mempalace_search or
mempalace_kg_query to verify.
Verify facts: If a memory seems outdated, check the knowledge graph
with as_of date filtering. Facts may have been invalidated.
Record learnings: After significant sessions, use mempalace_diary_write
or mempalace_add_drawer to persist new knowledge.
Invalidate, don't delete: When facts change, use mempalace_kg_invalidate
to end validity rather than deleting. This preserves the temporal record.
| Layer | Content | When Loaded | Token Budget |
|---|---|---|---|
| L0 | Identity context | Always (wake-up) | ~50 tokens |
| L1 | Critical facts | Always (wake-up) | ~120 tokens |
| L2 | Room recall | On-demand (scoped search) | Variable |
| L3 | Deep semantic search | On-demand (cross-palace) | Variable |
mempalace_status (~170 tokens total)mempalace_search with wing/room filtersmempalace_search| Need | Tool | Why |
|---|---|---|
| "What do we know about X?" | mempalace_search | Broad semantic search |
| "What's in the auth wing?" | mempalace_search with wing param | Scoped, higher precision |
| "What decisions did we make?" | mempalace_search (filter results client-side by hall type) | No hall filter exists on the tool |
| "Who works on what?" | mempalace_kg_query | Entity relationships |
| "What changed since January?" | mempalace_kg_timeline | Temporal history |
| "How are X and Y connected?" | mempalace_find_tunnels | Cross-wing discovery |
| "Save this decision" | mempalace_add_drawer | Verbatim storage |
| "Update a fact" | mempalace_kg_invalidate + mempalace_kg_add | Temporal fact management |
When adding content to the palace:
mempalace_check_duplicate before add_drawersource_file parameter when availableExpert guidance for creating Claude Code skills and agents. Use when working with SKILL.md files, authoring new skills, creating slash commands, or designing agent workflows.
Run 1-3 parallel quick-audit passes (code quality, security, error handling) over a diff and gate on findings. Use when smart-submit or gt-amend need to audit uncommitted changes before committing.
Decompose a feature into stacked PRs, ordered by dependency (plan-only). Use when breaking a feature into reviewable stacked PRs.