A curated collection of skill files for AI coding agents - Pi, OpenCode, Claude Code, and any agent that supports the .claude/<name>/SKILL.md convention - to help with Wikipedia, Wikimedia, and Wikidata-related tasks.
Whether you're editing a Wikipedia article, researching a topic, or building a Toolforge tool, these skills give your AI assistant expert guidance on Wikimedia's systems, policies, and data.
Each skill is a self-contained set of instructions, policy knowledge, code examples, and executable tooling that an agent loads on demand. This means agents get expert-level guidance without bloating their system prompt.
These skills are not just reference docs — they measurably improve AI agent performance on Wikimedia tasks. In controlled benchmarks comparing agents with and without skills loaded:
- 2.8× average speedup — agents with skills completed tasks faster (2.2× for simple tasks → 3.1× for complex cross-domain tasks)
- 50% silent failure rate → 0% — without skills, agents produced plausible-looking but wrong answers in half of all tasks. With skills, this dropped to zero.
- The gap widens with complexity — the most complex, cross-domain tasks showed the largest advantage. Skills help most where the agent needs to navigate multiple Wikimedia systems (API + policy + wikitext + Wikidata) in a single task.
In short: skills make agents faster, cheaper (fewer tokens wasted on wrong paths), and dramatically more reliable.
Most language editions of Wikipedia and other Wikimedia projects have community policies that restrict or ban the direct posting of content generated by large language models (LLMs). This includes articles, sections, discussion comments, or any other substantial prose drafted by AI assistants, as well as images and other media files, whether uploaded to Wikimedia Commons or directly to a single Wikipedia edition. Before using AI to generate content, you must consult all relevant AI policies. These are summarised on Meta-Wiki at AI policies by project. You must read the full policies and guidelines, not just the summaries.
These skills are designed to help you research, understand, and prepare content - not to produce final text for direct submission. If you use the biography-writing skill to draft an article, treat the output as a research aid and starting point, not a finished product. Any Wikipedia edit you make remains your own responsibility and must comply with Wikipedia's policies on verifiability, neutral point of view, and no original research.
Do not copy-paste AI-generated text directly into Wikipedia. Always rewrite, verify every claim against reliable sources, and ensure the result reflects your own editorial judgment.
This repository contains 58 skills organized into two groups.
💡 Explore the interactive skills network → — see how all 58 skills connect to each other in a force-directed graph. Drag nodes, zoom, and hover to discover cross-references between skills.
Skills are arranged from foundational (understanding how Wikipedia works) to specialized (analyzing and creating content).
| Skill | What it helps with |
|---|---|
| ▸ Understanding Wikipedia | |
| wikipedia-page-anatomy | Navigate and understand the structure of a Wikipedia article — infoboxes, categories, references, templates, navboxes, redirects, disambiguation, and protection levels |
| wikipedia-templates | Create, design, and understand MediaWiki templates — template syntax, parser functions, magic words, transclusion vs substitution, Lua modules, template types taxonomy, API detection, and maintenance workflows |
| wikipedia-categories | Understand and work with Wikipedia's category system — category trees, the three tests for valid categories (Verifiable/Neutral/Defining), topic vs. set categories, sort keys and DEFAULTSORT, API query patterns, Pywikibot generators, PetScan, overcategorization rules, and category maintenance |
| wikipedia-edit-history | Read and analyze Wikipedia's page revision history — diffs, edit summaries, user contributions, byte changes, rollback, and understanding how a page evolved over time |
| wikimedia-diffs | Fetch, compare, and interpret diffs between Wikipedia page revisions — wikitext changes, visual differences, and diff statistics via the Action API and REST API |
| wikiwho | Token-level authorship attribution via the WikiWho API — who wrote, removed, or reinserted each word, with content-persistence and editor attribution analysis |
| wikipedia-talk-page | Navigate and participate in Wikipedia talk pages — modern DiscussionTools (Reply Tool, Topic Subscriptions, Permalinks), Usability Improvements, community gadgets, wikitext conventions, WikiProject banners, archives, and talk page etiquette |
| wikipedia-wikiprojects | Understand and work with English Wikipedia's WikiProject system — finding relevant projects, interpreting assessment tables, using Popular pages and work lists, and navigating project directories |
| ▸ Working with Content | |
| wikimedia-wikitext | Parse, extract, and manipulate Wikipedia and MediaWiki wikitext (templates, infoboxes, citations, links) using proper AST-based tooling instead of fragile regex patterns |
| wikidata | Understand and query Wikidata — the free, collaborative, multilingual knowledge graph that underpins Wikipedia's inter-language links, Commons structured data, and semantic facts across all Wikimedia projects. Covers SPARQL, the Wikibase REST/Action APIs, RDF data dumps, and semantic web concepts |
| wikiportraits-event-series | Create Wikidata edition items and Commons year categories for recurring events — P155/P156 follows/followed-by chains, cancelled-year numbering, edition claims, and the category scheme |
| wikidata-vector-search | Query Wikidata by meaning, concept, or natural-language description using semantic embeddings — fuzzy semantic search, concept matching, similarity lookups, cross-lingual retrieval, and "find like this" when you don't know the exact QID or label |
| quickstatements | Build and run QuickStatements batches for Wikidata and Commons — the V1 command grammar (statements, qualifiers, references, ranks, item creation), value formatting, multilingual labels/descriptions/aliases, QS 2.0 vs 3.0, and a bundled validator/generator script |
| wikipedia-citations | Master Wikipedia citations — CS1/CS2 templates, Wayback Machine archiving, dead link detection, bare URL expansion, citation maintenance, and reference validation |
| wikipedia-wikitables | Create, parse, style, and fix MediaWiki wikitable syntax — delimiters, header/data cells, CSS classes, inline styling, rowspan/colspan, accessibility attributes, sortable and collapsible tables, and programmatic table generation from data |
| wikimedia-page-styling | Use TemplateStyles to load custom CSS on wiki pages — build responsive grid/flexbox layouts, card-based tile systems, color themes, button systems, and full visual design systems that transform plain MediaWiki pages into rich, interactive-looking interfaces |
| mediawiki-page-navigation | Build navigation systems in MediaWiki — menu bars, subpage hierarchies, breadcrumbs, tabs, and the template logic that powers them |
| mediawiki-translate-extension | Work with the Translate extension for multilingual wiki content — marking pages for translation, writing translatable templates, and building i18n-aware navigation |
| mint | Translate text and rich content via MinT — the Wikimedia machine translation service: 200+ languages, open NMT models (NLLB-200, OpusMT, IndicTrans2), plain/HTML/JSON/markdown formats |
| ▸ Commons & Media | |
| wikimedia-commons | Search, upload, and understand Wikimedia Commons — the free media repository of images, video, sound, 3D files, PDFs, and other media used across Wikipedia and its sister projects. Browse categories, find reusable media, and retrieve file metadata |
| commons-file-resolution | Resolve Wikimedia Commons file references to browser-usable HTTP URLs — direct origin URLs, thumbnails, Special:FilePath redirects, cache-busting with timestamps, Action API imageinfo queries, and CORS-aware serving patterns for web applications |
| wikimedia-commons-thumbnails | Generate, construct, and retrieve raster thumbnail previews for any Wikimedia Commons file — thumb URL scheme, iiurlwidth/iiurlheight API pattern, thumbmime format conversion matrix, responsive/retina URLs, REST API thumbnails, and SPARQL thumbnailUrl |
| wikimedia-commons-svg | Work with SVG files on Wikimedia Commons — viewing and retrieving raw SVG source vs PNG preview, W3C validation badges, creation tools (Inkscape, Illustrator), optimization (scour, SVGO), versioning and diffing SVG revisions, and SVG-specific templates and categories |
| wikimedia-commons-pdf | Work with PDF and DjVu documents on Wikimedia Commons — multi-page document model, page selection for thumbnails, page dimensions, uploading large documents, Wikisource proofread integration, OCR text extraction, and document metadata |
| wikimedia-commons-audio-video | Work with audio and video files on Wikimedia Commons — format policies and patent restrictions, uploading and transcoding, metadata via the Action API, keyframe thumbnails for video, the TimedMediaHandler player widget, TimedText subtitles, and creating derivative clips |
| wikimedia-commons-sdc | Add, edit, and manage Structured Data on Commons (SDC) — MediaInfo captions, depicts statements, copyright and license metadata, qualifiers, references, and batch/GLAM workflows via the Wikibase Action API, web UI, and community tooling |
| wikimedia-commons-sparql | Query Wikimedia Commons structured data via SPARQL — MediaInfo entities (M IDs), the depicts/copyright/licensing graph, Schema.org media metadata, and federated queries with Wikidata. Covers both the official WCQS endpoint and the QLever third-party endpoint |
| wikimedia-commons-categories | Create and disambiguate Wikimedia Commons categories from Wikidata data — the occupation-from-country pattern, by name people categories, definite-article country phrases, pluralization, existence probing, and homonym disambiguation, distilled from the production Catapult gadget |
| pattypan | Build pattypan upload spreadsheets (.xls) for batch Wikimedia Commons uploads — the two-sheet Data/Template format, path/name headers, template-variable matching, filename validation, and a bundled generator/validator script |
| flickr | Fetch photos from Flickr's read-only REST API (photosets, search, metadata, tags, geo) and prepare pattypan upload manifests for batch Wikimedia Commons uploads - license filtering, flickr2commons-style descriptions, and attribution |
| flickr-wayback-recovery | Recover a deleted or offline Flickr account from the Wayback Machine (CDX enumeration, metadata scraping, image download) and batch-upload the photos missing from Wikimedia Commons via pattypan — matched by Flickr ID |
| ▸ Analyzing Articles | |
| wikimedia-pageviews | Retrieve traffic and popularity statistics for Wikipedia articles using cached SQL properties (sorting/filtering) or the REST API (precise historical data) |
| xtools | Query XTools — the canonical Wikimedia statistics API: page info, top editors, edit counts, prose stats, and admin/patroller metrics for any wiki |
| wikimedia-media-usage-metrics | Measure and count the use of Wikimedia media files — transfers (mediacounts/mediarequests), embeds (GlobalUsage), reach (pageviews/CIM), external reuse — with verified gotchas, a decision tree, and a live report pipeline |
| wikimedia-page-assessment | Query Wikipedia article quality (FA/GA/B/C/Start/Stub) and importance ratings from WikiProject assessment banners on any Wikimedia wiki with the PageAssessments extension |
| wikimedia-ml-services | Score article quality, revert risk, edit quality (goodfaith/damaging), readability, topic classification, reference quality, language identification, content translation recommendations, article descriptions, and article country using Wikimedia ML inference APIs (Lift Wing and legacy ORES) |
| wikipedia-reference-verifiability | Analyze whether a Wikipedia page's references contain URLs — detect bare plain-text citations, template-based citations without url= parameters, shortened footnotes, and named ref reuse. Useful for article quality assessment, NPP triage, and citation maintenance |
| ▸ Writing & Improving | |
| wikipedia-en-biography-writing | Draft and edit English Wikipedia biographies following NPOV, verifiability, no original research, and biographies of living persons (BLP) policies |
| wikipedia-en-article-audit | Audit an English Wikipedia article for structural issues, factual errors, and NPOV violations, then produce a machine-readable task graph (DAG) that another agent can execute to fix all identified problems |
| wikipedia-notability-assessment | Evaluate whether a subject meets Wikipedia notability guidelines — the General Notability Guideline (GNG), all 13 subject-specific SNGs with decision trees, source quality evaluation, structured report generation, AfD-ready summaries, and common invalid arguments |
| ▸ Other Wikimedia Projects | |
| wikimedia-url-shortener | Create and expand w.wiki short URLs (Wikimedia's URL shortener) — 301 redirect behavior, the missing expansion API, the browser CORS trap, server-side resolution patterns, and creation via the meta.wikimedia.org action=shortenurl API |
| wiktionary-and-wikisource | Work with Wiktionary (dictionary entries, translation tables, etymologies, audio pronunciations, lexemes) and Wikisource (proofread page workflow, OCR text extraction, quality validation, compiled works) — the two largest Wikimedia content projects after Wikipedia |
| wikivoyage | Work with Wikivoyage — the free, collaborative, multilingual travel guide. Covers listing templates (See/Do/Buy/Eat/Drink/Sleep), the geographical hierarchy (continent → country → region → city), dynamic OSM maps via Kartographer, article status (stub/outline/usable/guide/star), the mobile-friendly image policy, Wikidata integration for coordinates, and batch operations |
Skills for building bots, web apps, and data pipelines on Wikimedia infrastructure.
| Skill | What it helps with |
|---|---|
| wikimedia-api-access | Access Wikipedia and Wikimedia APIs (REST, Action, SPARQL) with correct User-Agent headers, rate limiting, and 429/403 error handling |
| wikimedia-api-strategy | Choose the right Wikimedia API or tool for the task — decision framework covering REST API, Action API, SPARQL, SQL replicas, EventStreams, and Pywikibot, with latency/complexity/authentication trade-offs |
| wikimedia-auth-oauth | Authenticate Wikimedia API clients for editing, patrol, upload, and user-specific operations — OAuth 1.0a/2.0 flows, bot passwords, CSRF tokens, and permission checks |
| wikipedia-error-handling | Handle HTTP errors, rate limits, and API failures across all Wikimedia services — retry strategies, backoff patterns, error response formats, and recovery procedures for the Action API, REST API, SPARQL, Lift Wing ML, and EventStreams |
| wikimedia-security-and-privacy | Build tools that respect Wikimedia user privacy and security — data minimization, suppressed/deleted revision handling, deanonymization risks, AbuseFilter and block awareness, XSS prevention, and data retention policies for Toolforge tools |
| wikimedia-database | Execute SQL queries against Wikipedia database replicas (enwiki, wikidata, commonswiki) via an SSH tunnel to Toolforge, with connection management and data handling guardrails |
| wikimedia-toolforge | Manage Toolforge accounts, web services, Kubernetes pods, cron jobs, file deployment, and privacy-preserving CDN assets for Wikimedia tools |
| toolforge-nodejs | Deploy and manage Node.js web services on Wikimedia Toolforge Kubernetes — zero-dependency server patterns, webservice commands, PORT configuration, static file serving with caching headers, npm on NFS, environment variables, logging, and common pitfalls |
| toolforge-python | Deploy and manage Python web services on Wikimedia Toolforge Kubernetes — Flask and gunicorn setup, webservice commands, virtual environments on NFS, pip caching, PORT configuration, static file serving, logging, and cron jobs |
| wikimedia-phabricator | Navigate Wikimedia's Phabricator instance — search tasks, interpret task status and project tags, file bug reports, and track WMF development priorities |
| wikimedia-i18n-l10n-for-tools | Design multilingual Toolforge tools — message files and ICU plurals, language detection and fallback chains, RTL/bidi layout, Unicode normalization, cross-wiki domain mapping, batch Wikidata label fetching, and avoiding English Wikipedia assumptions |
| wikimedia-codex | Build Wikimedia-consistent web interfaces with the Codex design system — Vue 3 components, CSS-only components, design tokens, icons, and composables. For Toolforge tools, MediaWiki gadgets, and Wikimedia-adjacent frontends |
| ▸ Search & Data | |
| wikimedia-search-cirrussearch | Search Wikimedia wikis using CirrusSearch — syntax cheat sheet (insource, hastemplate, linksto, deepcategory, haswbstatement), API parameters, prefix vs full-text vs title search, ranking caveats, maintenance queries, and combining search with PetScan, SPARQL, and categories |
| wikimedia-petscan | Query Wikimedia projects with PetScan — multi-source category intersections, template filtering, SPARQL integration, Wikidata item filters, bulk data export, and the PSID stable query ID system |
| wikimedia-pageviews | Retrieve traffic and popularity statistics for Wikipedia articles using cached SQL properties (sorting/filtering) or the REST API (precise historical data) |
| wikimedia-page-assessment | Query Wikipedia article quality (FA/GA/B/C/Start/Stub) and importance ratings from WikiProject assessment banners |
| wikimedia-ml-services | Score articles and revisions using Lift Wing ML models — revert risk, article quality, topic classification, readability, language identification, and more. Includes ORES migration guidance |
| ▸ Automation | |
| pywikibot | Use Pywikibot — the Python library and CLI tool suite for automating work on MediaWiki sites. Covers installation, the core object model, page generators, the bot framework, built-in scripts, and Wikidata/Commons integration |
| wikimedia-eventstreams | Consume real-time streams of Wikimedia events (edits, page creations, deletions, moves, log entries) via Server-Sent Events (SSE). Covers stream schemas, client libraries, filtering, historical replay, and building live dashboards |
| If you want to... | Use this skill |
|---|---|
| ▸ Understanding Wikipedia | |
| Understand the structure of a Wikipedia article (infobox, categories, templates, navboxes) | wikipedia-page-anatomy |
| Understand how MediaWiki templates work — syntax, parser functions, magic words, Lua modules | wikipedia-templates |
| Create, inspect, or design a new template for a Wikipedia article | wikipedia-templates |
| Understand the category system — valid categories, sort keys, tree structure, and querying | wikipedia-categories |
| Evaluate whether a proposed category is valid (Verifiable/Neutral/Defining test) | wikipedia-categories |
| Find all pages in a category or category tree via API | wikipedia-categories |
| Find the intersection of two categories | bash .claude/skills/wikipedia-categories/assets/category-intersect.py "Physicists" "German scientists" |
| Inspect all categories assigned to an article with metadata | python3 .claude/skills/wikipedia-categories/assets/category-inspector.py Albert_Einstein |
| Explore a category hierarchy interactively from the CLI | bash .claude/skills/wikipedia-categories/scripts/category-tree.sh Physics 3 pages |
| Analyze page revision history, diffs, user contributions, and detect vandalism | wikipedia-edit-history |
| Compare two revisions and understand what changed | wikimedia-diffs |
| Find out who wrote specific sentences or tokens in an article, and each editor's share | wikiwho |
| Measure content persistence — which edits survived, or which text keeps getting reverted | wikiwho |
| Participate in talk page discussions, sign comments, and find WikiProject assessments | wikipedia-talk-page |
| Find the right WikiProject for an article topic | wikipedia-wikiprojects |
| ▸ Working with Content | |
| Translate text or article content between languages (200+ languages, open NMT models) | mint |
| Translate rich content while preserving HTML/JSON/markdown markup | mint |
| Extract infobox data, citations, or internal links from an article's wikitext | wikimedia-wikitext |
| Query Wikidata with SPARQL — find items, traverse properties, filter by value | wikidata |
| Find Wikipedia articles that lack a Wikidata item | wikidata |
| Search Wikidata by meaning (not exact name) — fuzzy cross-lingual concept search | wikidata-vector-search |
| Set up event series: Wikidata edition items, P155/P156 chains, and Commons year categories | wikiportraits-event-series |
| Check if a URL is archived on the Wayback Machine | bash .claude/skills/wikipedia-citations/scripts/archive-check.sh https://example.com |
| Inspect a summary of all citations on a Wikipedia page | bash .claude/skills/wikipedia-citations/scripts/citation-inspector.sh Albert_Einstein |
| Expand a bare URL into a proper citation template | bash .claude/skills/wikipedia-citations/scripts/expand-bare-url.sh https://example.com |
| Auto-generate a full citation from a URL, DOI, or ISBN | bash .claude/skills/wikipedia-citations/scripts/citoid-expand.sh 10.7554/eLife.32259 |
| Check a Wikipedia page for dead links and suggest archives | bash .claude/skills/wikipedia-citations/scripts/check-dead-links.sh Albert_Einstein |
| Create a wikitable from Python data | python3 .claude/skills/wikipedia-wikitables/assets/wikitable_tools.py |
| ▸ Commons & Media | |
| Search Commons for freely licensed images, audio, video, and documents | wikimedia-commons |
| Generate thumbnail URLs for any Commons file at any size | wikimedia-commons-thumbnails |
| Create responsive/retina image URLs from Commons files | wikimedia-commons-thumbnails |
| Serve Commons images in a browser app (handle CORS issues) | commons-file-resolution |
| Query Commons structured data via SPARQL (find files by license, depicts, camera) | wikimedia-commons-sparql |
| Add depicts/captions/copyright statements to Commons files | wikimedia-commons-sdc |
| Build a pattypan .xls spreadsheet for a batch Commons upload | python3 .claude/skills/pattypan/scripts/build_pattypan_spreadsheet.py --manifest files.csv --template template.wikitext --output upload.xls |
| Fetch Flickr photos into a pattypan manifest for a batch Commons upload | python3 .claude/skills/flickr/scripts/fetch_flickr.py --photoset <id> --user <nsid> --out files.csv then python3 .claude/skills/pattypan/scripts/build_pattypan_spreadsheet.py --manifest files.csv --template template.wikitext --output upload.xls |
| Recover photos from a deleted Flickr account via the Wayback Machine and upload them to Commons | flickr-wayback-recovery — python3 .claude/skills/flickr-wayback-recovery/scripts/cdx-photo-ids.py <nsid> <alias> then the rest of the pipeline |
| ▸ Analyzing Articles | |
| Get pageview statistics — how many people read an article? | wikimedia-pageviews |
| Get editor & page statistics — edit counts, top editors, article info, prose stats | xtools |
| See admin/patroller activity or batch page assessments on a wiki | xtools |
| Check article quality and WikiProject ratings (FA, GA, B, C, Start, Stub) | wikimedia-page-assessment |
| Score an article for quality and readability with ML | wikimedia-ml-services |
| Get a full article quality report (quality + readability + topics + reference risk) | python3 .claude/skills/wikimedia-ml-services/assets/article_quality_report.py Albert_Einstein en |
| Check if a Wikipedia page has URL-based references | wikipedia-reference-verifiability |
| ▸ Writing & Improving | |
| Draft a biography article with proper structure, citations, and NPOV | wikipedia-en-biography-writing |
| Audit an article for structural issues and generate a fix-it task plan | wikipedia-en-article-audit |
| Evaluate whether a subject is notable enough for Wikipedia | wikipedia-notability-assessment |
| Assess notability from the CLI | python3 .claude/skills/wikipedia-notability-assessment/assets/notability_checker.py "Jane Smith" --description "Professor at MIT, Nobel Prize winner" |
| Generate an AfD-ready notability summary | python3 .claude/skills/wikipedia-notability-assessment/assets/notability_checker.py "Test" --description "desc" --json |
| Parse a Wiktionary entry for definitions, translations, and pronunciations | wiktionary-and-wikisource |
| Check proofreading progress for a Wikisource work | bash .claude/skills/wiktionary-and-wikisource/scripts/ws-page-status.sh en "Index:Pride and Prejudice" |
| Extract listings from a Wikivoyage travel guide article | wikivoyage |
| Check the quality status of a Wikivoyage article (stub/outline/usable/guide/star) | wikivoyage |
| Extract all POI listings from a Wikivoyage article as JSON | bash .claude/skills/wikivoyage/scripts/extract-listings.sh "Tokyo" |
| ▸ Building Tools | |
| Call any Wikimedia API with correct User-Agent and rate limiting | wikimedia-api-access |
| Choose the right API for the task (REST vs Action vs SPARQL vs SQL vs EventStreams) | wikimedia-api-strategy |
| Authenticate tool users with OAuth or bot passwords for editing/uploading | wikimedia-auth-oauth |
| Handle API errors gracefully with retry and backoff | wikipedia-error-handling |
| Query Wikipedia SQL replicas from Toolforge | wikimedia-database |
| Deploy a web service, cron job, or bot on Toolforge | wikimedia-toolforge |
| Deploy a Node.js app on Toolforge | toolforge-nodejs |
| Deploy a Python app on Toolforge (Flask + gunicorn) | toolforge-python |
| Run Python bots at scale (edit 10K pages, harvest templates, transfer files) | pywikibot |
| Watch Wikipedia edits in real time and build live dashboards | wikimedia-eventstreams |
| Search wikis with advanced CirrusSearch syntax | wikimedia-search-cirrussearch |
| Run complex multi-source queries (categories + SPARQL + templates + Wikidata) | wikimedia-petscan |
| Build a multilingual Toolforge tool with i18n support | wikimedia-i18n-l10n-for-tools |
| Build custom wiki pages with CSS layouts (grid, flexbox, cards) | wikimedia-page-styling |
| Build a Wikimedia-consistent UI for your tool with the Codex design system | wikimedia-codex |
git clone https://github.com/fuzheado/Wikipedia-AI-Skills.git
cp -r Wikipedia-AI-Skills/.claude /path/to/your/project/cp -r Wikipedia-AI-Skills/.claude/skills/wikimedia-pageviews /path/to/your/project/.claude/skills/Once installed, your AI coding agent will discover the skill automatically through the skill tool. You can also open any SKILL.md file and paste its contents directly into an agent's instructions.
Pi discovers skills from anywhere on the filesystem via settings.json.
mkdir -p ~/.pi/repos
git clone https://github.com/fuzheado/Wikipedia-AI-Skills.git ~/.pi/repos/Wikipedia-AI-SkillsEdit ~/.pi/agent/settings.json:
{
"skills": [
"~/.pi/repos/Wikipedia-AI-Skills/.claude/skills"
],
"extensions": [
"~/.pi/repos/Wikipedia-AI-Skills/.pi/extensions/wikimedia-skills"
]
}The extension ships with a placeholder User-Agent. You must replace the contact information with your own — Wikimedia uses this to reach you if your requests cause issues.
Choose one of these methods:
Option A (recommended) — Persistent config that survives
git pull
Copy the template to your home config directory, then edit the copy:
# 1. Create the user config directory
mkdir -p ~/.config/wikimedia-skills
# 2. Copy the template (this is a one-time copy)
cp ~/.pi/repos/Wikipedia-AI-Skills/.pi/extensions/wikimedia-skills/config.json \
~/.config/wikimedia-skills/config.json
# 3. Open the copy and replace the userAgent value with your own info
$EDITOR ~/.config/wikimedia-skills/config.jsonIn that file, change the "userAgent" value to your own contact info. The format is:
<client>/<version> (<contact URL or email>) <project>
For example:
{
"userAgent": "MyBot/1.0 (https://github.com/yourname; yourname@example.com) SkillsDemo"
}Option B — Quick override (lasts one shell session)
export WIKIMEDIA_USER_AGENT="MyBot/1.0 (https://github.com/yourname; yourname@example.com) SkillsDemo"/reloadAfter reloading, the User-Agent injection extension is active: every curl, wget, python, and node command targeting a Wikimedia server automatically gets a compliant User-Agent — no 403 errors, no manual intervention needed.
The extension checks these locations in order (first match wins):
WIKIMEDIA_USER_AGENTenvironment variable~/.config/wikimedia-skills/config.json(your personal config — survivesgit pull)- The repo's
.pi/extensions/wikimedia-skills/config.json(shipped default — do not edit,git pulloverwrites it)
⚠️ Do not edit the repo'sconfig.jsondirectly. It is a template andgit pullwill overwrite your changes. Always use~/.config/wikimedia-skills/config.jsonor theWIKIMEDIA_USER_AGENTenvironment variable.
All shell and Python scripts in this repository follow a script compliance standard
defined in .claude/guidelines/script-audit-guidelines.md:
- ✅ Zero-argument guard — every script prints a helpful usage message when invoked with no arguments
- ✅ Portable bash — no bash 4+ features (
declare -A,${var,,}) that break on macOS's default bash 3.2 - ✅ Safe piped output —
curl | python3patterns use temp files with HTTP status checks instead of silently passing empty/invalid input - ✅
--helpsupport — every script responds to--helpwith full documentation - ✅ Deferred imports — Python scripts with optional dependencies show help before crashing on missing imports
See the full guidelines for the pre-commit hook, CI workflow templates, and checklist for new scripts.
Skills document many CLI commands, API calls, URLs, and code snippets. To guard
against hallucinated content — plausible-looking commands/modules/URLs that
do not exist (e.g. toolforge tools create, action=templatestyles,
prop=translationinfo) — this repo keeps ground-truth registries captured from
live systems, and every skill is verified against them in CI
(.github/workflows/skill-verification.yml):
| Check | Registry (generated) | Refresher |
|---|---|---|
CLI commands (toolforge, pwb.py, webservice, sql, become) |
scripts/command-registry.json |
scripts/refresh-command-registry.py (SSH to bastion + pywikibot repo) |
Action API modules (action=/prop=/list=/meta=) |
scripts/api-surface.json |
scripts/refresh-api-surface.py (live paraminfo, enwiki+wikidata+commons) |
| External URLs (status checked) | scripts/url-registry.json |
scripts/refresh-url-registry.py (incremental HEAD sweep, freshness rotation) |
Links/depends_on/wikilinks |
— (offline) | scripts/verify-links.py |
| Code snippet syntax (python/bash/json/js) | — (offline) | scripts/verify-snippets.py |
last_verified freshness |
— (offline) | scripts/verify-freshness.py |
Run all checks locally:
python3 scripts/verify-commands.py
python3 scripts/verify-api.py
python3 scripts/verify-links.py
python3 scripts/verify-snippets.py
python3 scripts/verify-freshness.pyRefresh the registries when the underlying systems change (or every few months):
TOOLFORGE_USER=<your-ldap-username> python3 scripts/refresh-command-registry.py
python3 scripts/refresh-api-surface.py
python3 scripts/refresh-url-registry.py # incremental: new + stale (>90d) + broken (>30d)
python3 scripts/refresh-url-registry.py --new-only # fast pre-push check after adding URLs
python3 scripts/refresh-url-registry.py --full # rare full audit of every URLURL refresh is incremental by default: a URL is only live-checked if it is new, outside the
~3-month freshness window (--max-age-days), or a known-broken URL due for a retry
(--retry-after-days). Known-good URLs are skipped, so normal runs check only the handful
that rotated in — no more multi-minute sweeps on every change. CI's verify-links.py is fully
offline; a scheduled workflow (.github/workflows/url-registry-refresh.yml, weekly) keeps the
registry fresh automatically.
Prose that documents removed commands ("the toolforge tools... family was
removed"), error demos, illustrative placeholder URLs, and POST-only endpoints
is automatically recognized and skipped.
New contributions should include tests. The project uses pytest with a test suite in
tests/ covering:
- YAML frontmatter validation - every SKILL.md must have valid
name,description,license, andcompatibilityfields - Content accuracy - key SOP claims, code examples, and guardrails are verified against the actual markdown
- Code execution - runnable Python assets (like the cross-API pipeline script and the Pywikibot quick reference) have mock-based unit tests
The skills in this repository were validated through a systematic A/B test — 24 parallel agent runs across 12 tasks in 3 rounds of escalating complexity — comparing agent performance with and without skill injection.
Key findings:
- 2.8× average speedup with skills loaded (2.2× for simple tasks → 3.1× for complex cross-domain tasks)
- 50% silent failure rate without skills vs. 0% with skills — the no-skills agents produced plausible-looking wrong answers in half of all tasks
- The gap widens with task complexity: the most complex tasks showed the largest advantage
See research/ab-testing/ for the full reports, including:
- Round-by-round results (R1: simple APIs, R2: multi-step workflows, R3: cross-domain orchestration)
- Meta-analysis synthesizing all 12 comparisons
- Skills improvement recommendations derived from test findings
- Code example consistency audit
pip install pytest pyyaml requests
python -m pytest tests/Every skill lives in .claude/skills/<name>/ with this structure:
my-skill/
├── SKILL.md # Agent-facing instructions with SOPs
├── scripts/ # Executable helpers (bash, python, etc.)
├── references/ # Deep reference docs loaded on-demand
└── assets/ # Templates, config files, sample data
The SKILL.md file includes YAML frontmatter for agent discovery:
---
name: skill-name
description: Short description for agent discovery
license: MIT
compatibility: opencode
---
## SOP: ...
...agent-facing instructions...For background on the design motivations and community tools that inspired this project, see PAWS, Quarry, and the Wikimedia AI Facility — a presentation covering the origins of PAWS (hosted Jupyter notebooks for Wikimedia) and Quarry (public SQL query interface for wiki databases), and what they suggest about designing an AI-assisted editing facility for the Wikimedia community.
See CONTRIBUTING.md for detailed guidelines on adding new skills, improving existing ones, and the pull request process. All skills must follow the YAML frontmatter format and pass the content accuracy checklist.
For the rationale behind skill sizing, tier assignments, and the split-vs-merge philosophy, see docs/design-philosophy.md.
MIT