wiki-summary

Wiki Summary Skill

The model already knows what Wikipedia said at training time; this skill fetches what it says now — which matters for anything living: people's roles, company facts, ongoing events, populations, "current CEO" questions. Wikipedia's REST API serves a clean summary per article over keyless HTTPS. The skill's discipline is attribution: the fetched extract is Wikipedia's voice, dated today; anything the model adds around it gets labeled as such.

What This Skill Produces

• The live extract — Wikipedia's current summary paragraph(s) for the topic, quoted as the source
• The link and metadata — canonical URL, and the description line ("American computer scientist")
• Disambiguation handling — when the title is ambiguous, the options, not a guess
• The command — exact curl, rerunnable

Required Inputs

Ask for these if not provided:

• The topic — resolved to an article title (spaces → underscores); ambiguous names get the disambiguation treatment, not a silent pick
• Language edition — en default; the endpoint pattern works on any edition (de.wikipedia.org, ja.wikipedia.org) and the user's question may belong in one
• Why they're asking — a fact-check wants the specific claim compared; a primer wants the extract; "has this changed" wants fetched-vs-recalled differences called out

Framework: The Endpoint and the Attribution Rules

1. The call: curl -s "https://en.wikipedia.org/api/rest_v1/page/summary/Alan_Turing" → JSON: title, description, extract (the summary text), content_urls.desktop.page (canonical link), type. URL-encode the title; spaces become underscores.
2. Search first when the title is uncertain: curl -s "https://en.wikipedia.org/w/rest.php/v1/search/title?q=turing&limit=5" → candidate titles. A type: "disambiguation" response means list the options and ask — a confident summary of the wrong John Smith is worse than a question.
3. Attribution is the product: the extract is quoted or clearly framed as "Wikipedia currently says…" with the link. Model elaboration goes outside that frame, labeled. A fact-check answer states: the claim, what the live article says, and whether they match.
4. Freshness honesty both directions: fetched beats recalled for living facts — but Wikipedia itself lags and errs; for high-stakes facts the answer notes it's one (good) source, and breaking-news topics may be mid-edit.
5. The changed-since-training move: when the fetched extract contradicts what the model would have said, say so explicitly — "training-era memory said X; the live article now says Y" — that delta is often exactly what the user was probing for.

Output Format

[Article title] — [description line]

> [The live extract, as Wikipedia's voice]

[Fact-check mode: the claim vs. the extract, verdict stated]

[Model additions, if any, under a labeled line]

Source: [canonical article URL] · fetched [date] · rerun: [exact curl]

[Disambiguation case: the candidate list and the ask]

Quality Checks

• [ ] The extract is attributed to Wikipedia and dated — never blended into model voice
• [ ] Ambiguous titles produced options, not a guess
• [ ] Fact-checks compare the specific claim to the specific sentence
• [ ] Training-memory vs. live-article deltas are called out when found
• [ ] The canonical URL appears

Anti-Patterns

• [ ] Do not paraphrase the live extract into model voice — the fetch's value is the attribution
• [ ] Do not silently pick among namesakes — disambiguate out loud
• [ ] Do not treat Wikipedia as final authority for high-stakes facts — one good source, framed as such
• [ ] Do not answer "what does Wikipedia say" from memory — that question is a fetch instruction by definition
• [ ] Do not skip the URL — the link is the receipt