Wiki Conflicts (Dexio)
Handle contradictions and outdated claims in an LLM wiki without silently overwriting: compare dates, provenance and the current system of record, replace only when one side is cle…
Dexio
@dexio
Install
$ openclaw skills install @dexio/dexio-wiki-conflictsResolve conflicts in a wiki
The costliest wiki error is not a missing fact but a wrong one that looks settled. It usually arrives as a quiet overwrite: a newer source, or a newer agent, replaces a claim without checking whether the old one was better sourced. The next agent cites the new line, and the error compounds.
Decide which case you are in
- Superseded. The old claim was true then and something changed: a price moved, a decision was reversed, a person changed roles. Replace the claim, keep a dated line of what it replaced when the history matters ("Was $8 until 2026-09-01"), and cite the new source.
- Corrected. The old claim was wrong. Replace it, cite the evidence, and say in the change note that it was a correction and what was wrong. Search the wiki for other pages that repeated the wrong claim and fix them too.
- Unresolved. Two sources disagree and neither is clearly authoritative. Keep both (format below), mark the page, and escalate if it matters.
- Not a conflict. Different scopes, dates or definitions ("users" versus "paying users"). Make the difference explicit on the page so the next reader does not see a conflict.
How to judge authority
In rough order:
- The system of record, checked now: the live config, the contract, the account, the person who decided.
- A primary source: the organization's own filing, docs or data, dated.
- A careful secondary source that shows its method.
- Vendor claims about themselves, self-reports, and aggregators.
- Anything recalled rather than read.
Newer beats older only within the same tier. A dated primary source beats a newer blog post that does not cite one.
Recording an unresolved conflict
On the page, next to the claim:
Monthly active users: disputed.
- 24,700 (company blog, 2026-08-14, vendor-sourced)
- about 9,000 (third-party panel estimate, 2026-09-02; counts web only)
Not resolved as of 2026-09-28; the panel cannot see API usage.
In the frontmatter, set contested: true and lower confidence if the page carries one.
When it is resolved, remove the flag, keep the winning claim with its source, and say in the
note what settled it.
Escalate
Escalate to a person, rather than choosing, when the conflict bears on a decision, money, a commitment to someone outside, or a claim other pages rely on. Say what conflicts, the sources on each side, which you think is right and why, and what would settle it.
Stale claims
A claim can be wrong without any contradicting source: it was simply true a year ago. When
you find one (from wiki-lint's stale list or while reading), recheck it at the source.
If it holds, bump updated and note that you rechecked it. If you cannot recheck it now,
add a stale_after date or an inline "as of" date so readers know how old it is.
Pitfalls
- Resolving by recency alone.
- Deleting the losing claim without a trace. History keeps it, but the next reader will not look at history; one dated line saves them re-adding it.
- Fixing one page and leaving the claim on the three pages that copied it. Search for it.
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.