一键导入
search-wiki
Use the committed-wiki MCP tools (or bash CLI as fallback) to search pages, traverse wiki relations, and bridge back to the corpus.
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。
菜单
Use the committed-wiki MCP tools (or bash CLI as fallback) to search pages, traverse wiki relations, and bridge back to the corpus.
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。
基于 SOC 职业分类
| name | search-wiki |
| description | Use the committed-wiki MCP tools (or bash CLI as fallback) to search pages, traverse wiki relations, and bridge back to the corpus. |
| allowed-tools | Bash(wikify wiki *) Bash(wikify data *) mcp__wikify__context_set mcp__wikify__context_show mcp__wikify__wiki_schema mcp__wikify__wiki_find mcp__wikify__wiki_show mcp__wikify__wiki_traverse mcp__wikify__corpus_show |
Use this skill to inspect the committed wiki. It teaches the read/query surface; it does not decide whether to refine, expand, or stop.
The wikify MCP server (mcp__wikify__wiki_*) is the canonical
read surface. The MCP session keeps the wiki index loaded across
calls. Bash wikify wiki * re-loads indexes per invocation. Use bash
only when the MCP server is not bound, when debugging the underlying
CLI, or when piping handles into shell tools.
Bind the bundle once per session:
mcp__wikify__context_set(bundle_path="<bundle>")
mcp__wikify__context_show()
wiki_find() (empty query lists all).wiki_show(handle="<title-or-slug>") returns the page; pass
full=True for the body. Handles accept the natural title or file
slug, matched case- and separator-insensitively (so
"Atomic Layer Deposition" resolves whether the file is
Atomic Layer Deposition.md or atomic-layer-deposition.md), or a
unique case-insensitive prefix.wiki_find(query="X", mode="hybrid").
Modes: hybrid (default), bm25, semantic, text.wiki_traverse(handle="<slug>", to="<relation>") walks one wiki
hop: links (outgoing), linked-by (incoming), co-evidence
(pages sharing source docs), similar, see-also,
category / categories, and evidence (emits chunk: handles
for the corpus).children, parent, pages.wiki_traverse(..., to="evidence") handles into
corpus_show(handle=...).The bash CLI also exposes wikify wiki repl --run <bundle> for
iterative committed-page search; MCP achieves the same with a bound
context.
kind=data artifacts live in a separate store from the wiki page
graph. They render and appear in navigation, but the wiki tools never
see them: wiki_find / wiki_show / wiki_traverse (and their bash
equivalents) return error="page_not_found" for a data table. That
result is expected, not a missing page. Do not retry on the wiki side.
Fall back to the data noun instead:
wikify data list --run <bundle> # list stored data claims
wikify data show <claim_id> --run <bundle> # show one claim
wikify data query --run <bundle> ... # filter the claim store
wikify data list-artifacts --run <bundle> # list kind=data tables
mcp__wikify__wiki_find(query="") # list all
mcp__wikify__wiki_find(query="ALD vs CVD", top_k=5)
mcp__wikify__wiki_find(query="resistive switching", mode="semantic")
mcp__wikify__wiki_find(query="atomic layer deposition", mode="text")
mcp__wikify__wiki_show(handle="Atomic Layer Deposition", full=True)
mcp__wikify__wiki_traverse(handle="Atomic Layer Deposition",
to="links", top_k=10)
mcp__wikify__wiki_traverse(handle="Atomic Layer Deposition",
to="categories")
mcp__wikify__wiki_traverse(handle="category:materials", to="pages")
mcp__wikify__wiki_traverse(handle="Atomic Layer Deposition",
to="evidence")
# -> for each returned chunk:
mcp__wikify__corpus_show(handle="<chunk-handle>")
wikify wiki list --run <bundle>
wikify wiki find "ALD vs CVD" --run <bundle> --top-k 5
wikify wiki find "resistive switching" --mode semantic --run <bundle>
wikify wiki find "atomic layer deposition" --run <bundle> --text
wikify wiki show "Atomic Layer Deposition" --run <bundle> --full
wikify wiki traverse "Atomic Layer Deposition" --to links \
--top-k 10 --run <bundle>
wikify wiki traverse "Atomic Layer Deposition" --to evidence \
--format quiet --run <bundle> \
| xargs -I {} wikify corpus show {} --corpus <corpus>
wikify wiki repl --run <bundle>
wikify wiki check --run <bundle>
references/wiki-cli-patterns.md - committed wiki command grammar and
use cases.references/wiki-recursive-search.md - recursive wiki search loops.references/wiki-corpus-bridges.md - moving from wiki evidence to
corpus search.../reference/references/cli/grammar.md - shared CLI grammar.../reference/references/cli/output-contract.md - output conventions.Researcher-style iterative wiki builder. Editor orchestrator dispatches explorer subagents that walk the corpus via named recursive patterns (P1-P5), gather evidence into notebook dossiers, and write pages when a composite maturity score crosses the gate. A DATA wave harvests verifiable numbers/tables into a claim store and consolidates them into evolving kind=data artifact tables. Coverage of the corpus chunk set is the primary objective. Re-entrant on the same bundle when new corpus material arrives.
Recursive exploration pattern library for wikify. Five named, depth-bounded procedures (P1 hub-anchor, P2 citation-walk, P3 semantic-boundary, P4 exact-term sweep, P5 gap-explorer) that compose existing corpus_find / corpus_traverse / corpus_citation_walk / wiki_find MCP primitives and append evidence into per-slug notebooks. Editor invokes one pattern per Task with a slug-disjoint target list.
Shared Wikify reference material for Claude Code skills. Use when working on Wikify skills, workflows, prompts, bundle state, writing rules, schemas, citation grounding, CLI grammar, exploration patterns, or MCP setup.
Produce a Wikify WriteResponse from supplied page context and evidence. Use when a workflow has selected a concept/page and needs encyclopedic article, person-page, comparison, or refinement prose grounded in evidence. Does not commit pages or decide readiness.
Consolidator that turns the bundle claim store into evolving data-artifact tables. Picks a dense subject-by-property theme, writes a durable view spec, and commits a kind=data wiki page that renders as an HTML table with per-cell references and re-derives from the spec as more data lands.
Extractor that harvests verifiable numeric/factual data points from corpus chunks, tables, and figure captions into the bundle claim store. Runs as a dedicated harvest pass over table-assets plus number-dense chunks, or piggybacks on chunks an explorer just read. Every point carries a verbatim grounding quote; the CLI gate rejects numbers it cannot locate in the source.