Wiki Query (Dexio)
Answer a question from an LLM wiki: find and read the relevant pages, answer with the page paths you relied on, check anything mutable at its original source, say plainly what the …
Dexio
@dexio
Install
$ openclaw skills install @dexio/dexio-wiki-queryAnswer from a wiki
A wiki earns its upkeep when answers come from it faster and better than from scratch, and when good answers are kept so the next question starts further ahead.
Steps
- Orient on the question (
wiki-orient): search two or three phrasings, read the canonical page and the neighbors that bear on it. - Answer from what you read, and cite it. Give the page paths the answer rests on. If you combine pages, say so. If a page's claim carries a trust label (vendor-sourced, estimate, unverified), carry the label into the answer.
- Check mutable facts at their source before presenting them as current: prices,
configuration, account state, who holds a role, anything with a date older than the
decision it will feed. Say what you checked. If the source disagrees with the wiki, the
answer uses the source and the wiki gets fixed (
wiki-recordorwiki-conflicts). - Say what the wiki does not cover. "The wiki has nothing on X" is a useful answer; a plausible guess presented as if it came from the wiki is a harmful one. If you fill the gap from outside research, mark which parts came from where.
- File the answer back when it is substantial. If the answer needed several pages, outside research, or a judgment worth keeping, write it as a query page (below). Trivial lookups are not filed.
- Fix what you found on the way. A broken link, a stale date, a missing link between two pages you needed together: fix it with a small edit and a note, or report it.
Query page
Under queries/ (or where the schema puts them), named for the question:
---
title: Which vendors include SSO on their base plan?
description: One line with the short answer, not the question.
created: YYYY-MM-DD
updated: YYYY-MM-DD
type: query
tags: [...]
sources: [pages and URLs consulted]
confidence: high | medium | low
---
# Which vendors include SSO on their base plan?
Question (who asked, date).
## Short answer
Two to four sentences.
## Evidence
What was consulted, with links to wiki pages and sources, and the numbers that carry the
answer.
## Confidence and gaps
What would change the answer, and what was not checked.
Link the query page from the pages it draws on when it adds something they lack, so the answer is found from the topic and not only by search. When a follow-up question arrives in the same thread, add it as a section of the same query page rather than a new page.
Pitfalls
- Answering from memory with the wiki open. If you did not read it in the wiki or a source this session, it is unverified; say so.
- Citing a page for a claim the page does not make. Quote or paraphrase closely enough that the reader can find the sentence.
- Filing every answer. A
queries/folder full of one-line lookups buries the real syntheses.
Top skills in this category
self-improving agent
@pskoettCaptures and maintains learnings, errors, and corrections. Use when a command fails, the user corrects an assumption, a capability is missing, knowledge is outdated, or a better approach is found. Review relevant saved guidance before major tasks, reuse, and promotion; revalidate it according to wha
Self-Improving + Proactive Agent
@ivangdavilaSelf-reflection + Self-criticism + Self-learning + Self-organizing memory. Agent evaluates its own work, catches mistakes, and improves permanently. Use when...
ontology
@oswalpalashTyped knowledge graph for structured agent memory and composable skills. Use when creating/querying entities (Person, Project, Task, Event, Document), linkin...
Humanizer
@biostartechnologyRemove signs of AI-generated writing from text. Use when editing or reviewing text to make it sound more natural and human-written. Based on Wikipedia's comprehensive "Signs of AI writing" guide. Detects and fixes patterns including: inflated symbolism, promotional language, superficial -ing analyses, vague attributions, em dash overuse, rule of three, AI vocabulary words, negative parallelisms, and excessive conjunctive phrases.
Obsidian
@steipeteWork with Obsidian vaults (plain Markdown notes) and automate via obsidian-cli.