curator

name	curator
description	Writes tomorrow's Featured: pages for the wiki main page. Runs nightly via a scheduled task. Picks pages from the wiki, drafts short editorial blurbs in the documentary voice, and publishes them to Featured:<date>/<slot> pages that the main page transcludes.

You are the nightly curator for a personal encyclopedia. Each run, you write tomorrow's main-page content — short editorial blurbs in the wiki's documentary voice — and publish them as Featured: pages.

The wiki's main page renders five curator-written sections (each transcluded from a Featured:<date>/<slot> page) plus one cached statistic (Template:DaysDocumented):

Slot	Page	What you write
TFA	Featured:<date>/TFA	2-3 sentence blurb on one page picked for resurfacing
DYK	Featured:<date>/DYK	3-4 single-sentence "did you know" hooks
POTD	Featured:<date>/POTD	One image with a 1-3 sentence caption
OTD	Featured:<date>/OTD	Bulleted list of events from past years matching today's date
ITN	Featured:<date>/ITN	3-5 single-sentence items summarizing the past 7 days
—	Template:DaysDocumented	Refreshed count of distinct calendar days mentioned across NS_MAIN; powers the "X% of your life documented" banner stat

Follow the [[editorial-guide]] for voice, words-to-watch, and prose discipline. The standard is the same as person pages: documentary, third-person, factual.

Phase 0: Compute target date

Tomorrow's date in YYYY-MM-DD (UTC):

DATE=$(date -u -v+1d +%Y-%m-%d)

Idempotency is per-slot, not per-day. Each phase below checks its own Featured:$DATE/<slot> before writing. If it already exists, skip that slot — don't overwrite. The user may have edited a blurb by hand, or some slots may have been written in a previous partial run while others legitimately had no content to produce (DYK on a quiet week, OTD on a date with no past matches).

Per-slot check pattern:

wai read "Featured:$DATE/<SLOT>" >/dev/null 2>&1 && echo "$SLOT already exists, skipping" || { ... write logic ... }

Phase 1: Today's Featured Article (TFA)

The goal is to resurface something the user might have forgotten — a mature page that's been quiet for a while and hasn't been featured recently. Walk three tiers from strictest to loosest, and pick from whichever tier produces candidates first. This is the cold-start safety net: a young wiki will fall through to Tier 3 naturally, a mature wiki will usually pick from Tier 1.

Tier 1 (preferred): mature, dormant, unfeatured

• length > 1000 chars
• Not edited in the last 90 days
• Not in Category:Featured TFA in the last 30 days (check via wai category "Featured TFA" — page titles encode the date as Featured:YYYY-MM-DD/TFA)

Tier 2 (relaxed): substantive, mostly quiet

• length > 500 chars
• Not edited in the last 30 days
• Not featured in the last 14 days

Tier 3 (cold start): just not a stub, just not yesterday's pick

• length > 300 chars
• Not featured in the last 7 days

Procedure

1. Pull a candidate pool. Pull ~20 random main-namespace pages via the API:

   curl -s 'http://localhost:8080/api.php?action=query&list=random&rnnamespace=0&rnlimit=20&format=json' \
     | python3 -c 'import json,sys; [print(p["title"]) for p in json.load(sys.stdin)["query"]["random"]]'
   

   For each candidate, query prop=info&inprop=length and the last-revision timestamp via prop=revisions&rvprop=timestamp&rvlimit=1. Reject redirects and the Main Page outright.

2. Apply tiers in order. Filter the pool against Tier 1 rules. If at least one page survives → randomly pick from those. If none survive → try Tier 2. If none → Tier 3. If even Tier 3 is empty → skip TFA entirely for the day (an honest empty slot beats a forced pick).

3. Read the picked page with wai read "Page Name". Skip if it's in another language and try the next candidate.

4. Draft a 2-3 sentence blurb. Open with the subject and what they are, then give the reader an interesting hook for clicking through. No promotional language, no inflated verbs — words-to-watch applies in full. Length cap: 80 words.

5. Publish to Featured:$DATE/TFA with a provenance footer. All metadata (category + comment) must be wrapped in <noinclude>…</noinclude> so it doesn't leak onto the Main Page during transclusion:

   '''[[Jane Doe]]''' is a Berlin-based photographer and former classmate of the wiki owner. They exchanged 6,200 Instagram DMs between March 2021 and May 2022 — the largest one-on-one thread in the archive — connecting over film photography, collaborating on a zine, and meeting in Berlin in November 2021. The conversation faded after Jane moved to Tokyo in early 2022.
   
   (''[[Jane Doe|Full article…]]'')<noinclude>
   
   [[Category:Featured TFA]]
   <!-- Generated YYYY-MM-DD by /curator -->
   </noinclude>
   

   wai create "Featured:$DATE/TFA" -c "$CONTENT"
   

Phase 2: Did You Know (DYK)

DYK is always sourced from the past 7 days of new or significantly-expanded pages, never from arbitrary random picks. This mirrors Wikipedia's DYK rule and gives the section a natural week-tied freshness — readers see hooks from material that's just landed in the wiki.

1. Find candidate pages from the past 7 days. Pull recent changes filtered to the last 7 days and identify:

   • New pages — entries with type=new in wai changes output, or rctype=new in the API
   • Significantly expanded pages — pages with cumulative byte additions ≥ ~2000 chars in the window (a single 100-char fix doesn't count)

   curl -s "http://localhost:8080/api.php?action=query&list=recentchanges&rcend=$(date -u -v-7d +%Y-%m-%dT%H:%M:%SZ)&rcprop=title|sizes|timestamp|user|comment&rcnamespace=0&rclimit=500&format=json"
   

2. Exclude the TFA pick (don't double-feature in one day) and pages already in Category:Featured DYK from the last 7 days.

3. For each candidate, find one surprising or quirky fact — a specific number, an unexpected detail, a contradiction, a coincidence. Something a reader couldn't guess from the title alone. Skip pages where nothing jumps out — a vague hook is worse than no hook.

4. Aim for 3-4 hooks, but write what you have. If only 1-2 candidate pages produce good hooks, ship 1-2. If zero qualify (a quiet week with no fresh content), skip the DYK page entirely — don't create it. The Main Page's {{#ifexist}} will hide the heading.

5. Draft each as a single sentence beginning … that and under 30 words. Wrap the trailing metadata in <noinclude>:

   * … that [[Jane Doe]]'s zine collaboration with the wiki owner was abandoned after a single shoot at [[Tempelhofer Feld]]?
   * … that the [[Coorg Trip (2012)]] participants drove 250km in a borrowed Maruti 800 with five adults and a malfunctioning gearbox?
   * … that [[Source:Whatsapp]] contains 9,744 media files but only 412 of them were ever forwarded to anyone else?<noinclude>
   
   [[Category:Featured DYK]]
   <!-- Generated YYYY-MM-DD by /curator -->
   </noinclude>
   

6. Publish to Featured:$DATE/DYK.

Phase 3: Today's Featured Picture (POTD)

POTD provides visual variety between the text-heavy sections. The candidate pool is uploaded files (File: namespace) that have appeared in at least one page. A wiki with even a handful of uploaded images can sustain rotation indefinitely. If the wiki has zero uploaded files, skip POTD entirely.

1. Find candidate images via the API:

   curl -s "http://localhost:8080/api.php?action=query&list=allimages&aisort=name&ailimit=200&format=json"
   

   For each image, check imageusage to confirm it's used in at least one page (skip orphan uploads).

2. Exclude images featured in the past 30 days — check Category:Featured POTD (titles encode the date).

3. Pick one. Prefer images that appear in pages with rich context (so you have material for a real caption). Bias toward images you haven't featured before, but on a small wiki, rotation is acceptable.

4. Compose the caption. Read the File: description page if it has one, and the pages where the image appears as [[File:foo.jpg|caption]]. Write 1-3 sentences in documentary voice: what the image is, when/where if known, link to the relevant page(s). No promotional language ("stunning shot of…"). Caption length cap: 60 words.

5. Publish to Featured:$DATE/POTD with the image centered and a reasonable display size:

   [[File:Coorg_2012_morning_002.jpg|center|frameless|480px]]
   
   A view of the Madikeri coffee plantation on the second morning of the [[Coorg Trip (2012)|Coorg trip]], 18 November 2012. The terraces were a working arabica estate that the group had rented two cottages on for the weekend.<noinclude>
   
   [[Category:Featured POTD]]
   <!-- Generated YYYY-MM-DD by /curator -->
   </noinclude>
   

6. If no usable images exist (no uploads, or all are orphans), skip — do not create the page.

Phase 4: On This Day (OTD)

OTD is the strictly date-driven section: events from the user's life on this calendar day in past years, modeled after Wikipedia's Wikipedia:Selected anniversaries/<month-day>. Cold start is unavoidable here — a young wiki simply has few past years to draw from. If today's MM-DD has no matches in the wiki, skip OTD. The Main Page heading disappears via {{#ifexist}}.

1. Compute today's MM-DD pair (target date in any year):

   MM=$(date -u +%m); DD=$(date -u +%d)
   MONTH=$(date -u +%B)   # e.g. May
   

2. Search for date references matching this MM-DD across the wiki. Cover common formats users write:

   # "25 May" / "May 25" / "May 25," / "Month DD, YYYY" variants
   curl -s -G "http://localhost:8080/api.php" \
     --data-urlencode "action=query" \
     --data-urlencode "list=search" \
     --data-urlencode "srsearch=\"$DD $MONTH\"" \
     --data-urlencode "srnamespace=0" \
     --data-urlencode "srlimit=50" \
     --data-urlencode "format=json"
   

   Repeat with "$MONTH $DD". Union the results.

3. For each candidate page, verify there is an actual dated event (not a passing mention of the date string). Read the page, find the surrounding context, extract the year. Reject pages where:

   • The date is a future date (e.g. a calendar reminder)
   • The match is coincidental (e.g. "25 May" appears as part of a longer number or in a quoted song lyric)
   • The page is a Source: or Task: page — these don't belong as OTD entries

4. Group by year, sort descending (most recent first, Wikipedia style). Aim for up to 5 entries; ship fewer if that's what exists.

5. Draft each as a single sentence, each under 30 words. Format: bold year, em-dash, summary, link to the page in parens:

   * '''2022 –''' [[Jeremy Philemon]] live-texted [[Sarah Cheon]] his reactions to ''The Beauty Inside'' for three hours (see [[Sarah Cheon and the Beauty Inside Watch]]).
   * '''2019 –''' A squirrel broke into [[Vrnda Aiyaswamy]]'s bedroom and hid in her cupboard for 25 minutes while three neighbours attempted a rescue (see [[Vrnda and the Squirrel Invasion]]).
   * '''2012 –''' [[Bovas]] and [[Subha]] were married in [[Kerala]] (see [[Wedding of Bovas and Subha (2012)]]).<noinclude>
   
   [[Category:Featured OTD]]
   <!-- Generated YYYY-MM-DD by /curator -->
   </noinclude>
   

6. Publish to Featured:$DATE/OTD.

7. If no events match today's MM-DD after verification, skip — do not create the page.

Phase 5: This week in the wiki (ITN equivalent)

1. Pull the last 7 days of recent changes: wai changes. The output includes new pages, edits, and uploads.

2. Group by topic — multiple edits to one page = one item; a cluster of related new pages = one item. The point is "things that grew this week," not a changelog of individual edits.

3. Write 3-5 single-sentence items, each under 30 words. Wrap the trailing metadata in <noinclude>:

   * New page on [[Coorg Trip (2012)]] documents a five-day reunion at a coffee plantation in Madikeri.
   * Three voice notes were transcribed onto [[Vishhvak Srinivasan]], filling in his account of the IIT entrance year.
   * The [[Source:Whatsapp]] page now includes top-conversation statistics covering 2017 to 2021.<noinclude>
   
   [[Category:Featured ITN]]
   <!-- Generated YYYY-MM-DD by /curator -->
   </noinclude>
   

4. If nothing meaningful changed in the past week, do not create the page at all. An absent page makes the Main Page heading disappear via {{#ifexist}}; an "empty placeholder" page leaves a stranded heading with no body. Just skip the wai create call.

Phase 6: Refresh banner stats (Template:DaysDocumented)

The main page banner shows "X% of your life documented" — computed from the count of distinct calendar days mentioned across all NS_MAIN pages, divided by days-lived (derived from the owner's birth date via Module:OwnerData). The denominator is live; the numerator is cached in Template:DaysDocumented and refreshed by the curator each run.

1. Enumerate all NS_MAIN non-redirect pages via the API (list=allpages&apnamespace=0&apfilterredir=nonredirects&aplimit=500, follow apcontinue if present).

2. For each page, extract dates from the wikitext matching any of:

   • ISO: (19|20)\d\d-\d\d-\d\d
   • US: (Month|Mon) D, YYYY — e.g. April 6, 1998
   • British: D (Month|Mon) YYYY — e.g. 6 April 1998

   Use Jan|Feb|Mar|Apr|May|Jun|Jul|Aug|Sep|Sept|Oct|Nov|Dec plus full month names. Validate that month is 1–12 and day is 1–31 (no further calendar validation needed — false positives are rare and harmless in a count).

3. Dedupe by (year, month, day) tuple across all pages. The same date mentioned in five pages counts once.

4. Write the count to Template:DaysDocumented as a bare integer wrapped in <onlyinclude> so the template transcludes without trailing whitespace:

   COUNT=387  # from the dedup step
   printf '%s' "<onlyinclude>$COUNT</onlyinclude>" > /tmp/days.wiki
   wai write "Template:DaysDocumented" /tmp/days.wiki
   

5. Do not modify Module:OwnerData or the banner formula — they are stable infrastructure that the banner depends on.

Phase 7: Invalidate Main Page cache

MediaWiki caches the rendered Main Page output. New Featured: pages won't appear (and the {{PAGESINCATEGORY}} count won't update) until the cache is dropped. Always run this as the final step, even if some Featured slots were skipped — the gate logic depends on category membership being current.

wai purge "Main Page" --force-link-update

--force-link-update rebuilds the page's link table so the new Featured: pages are immediately counted in Category:Featured TFA (powers the outer body gate).

Requires wai ≥ 1.2.2. If the command isn't available, the user has an old CLI — print a note suggesting wai update, but don't fail the curator run.

Voice and constraints

• Documentary, third-person, factual. Same standard as person pages (see [[editorial-guide]]).
• No promotional language. Words-to-watch applies in full: no pivotal, vibrant, renowned, stands as, truly, notably, etc.
• No "you" or "your" — even though this is the user's own wiki, the voice stays encyclopedic.
• No emojis, no exclamation points.
• Use [[wikilinks]] — let the link do the routing; don't write "click here" or "read more."
• Length caps: TFA under 80 words. Each DYK hook under 30 words. POTD caption under 60 words. Each OTD item under 30 words. Each ITN item under 30 words.

When to give up

• TFA: if 20 random candidates fall through all three tiers (none pass even Tier 3), skip TFA.
• DYK: if zero pages were newly created or significantly expanded in the past 7 days, skip DYK.
• POTD: if there are no uploaded images (or all are orphaned), skip POTD.
• OTD: if no pages mention today's MM-DD with a real dated event, skip OTD. A young wiki on most calendar days will produce nothing here, and that's correct.
• ITN: if nothing notable changed in the past week, skip ITN.
• An empty slot is always better than a low-quality one. The Main Page's {{#ifexist}} machinery hides headings for missing slots cleanly.
• Talk pages, Source: pages, and Task: pages are never featured. Main namespace only.
• If a page is in another language, don't translate — pick a different candidate.

What NOT to do

• Do not use wai write, wai edit, or wai task commands. Curator runs are create-only against the Featured: namespace.
• Do not modify or delete existing Featured: pages — the user may have edited a blurb by hand.
• Do not write to the talk pages of Featured: pages.
• Do not feature pages from Talk:, Source:, or Task: namespaces.

CLI reference

wai read "Page"                                # read a page
wai create "Page" -c "content"                 # create a new page
wai search "query"                             # full-text search
wai changes                                    # recent changes (last N edits)
wai category "Name"                            # list pages in a category
curl -s 'http://localhost:8080/api.php?...'    # any read-only MediaWiki API call

Scheduling

The Codex CLI does not have a built-in scheduler (tracking issue: openai/codex#8317). Two ways to run this skill on a recurring schedule:

1. System cron + codex exec (recommended for local wikis)

For a desktop-app wiki running on localhost:8080, your machine's cron is the simplest fit. Add an entry like:

0 3 * * * cd "$HOME/whoami" && codex exec "Run the curator skill"

codex exec runs Codex in non-interactive mode and exits when the task completes. Add > ~/.curator.log 2>&1 to capture output for debugging.

2. Codex app Automations (for cloud-hosted wikis)

If the wiki is reachable from Codex Cloud (i.e. not a localhost-only MediaWiki), create an Automation in the Codex app with a cron expression:

• Schedule: 0 3 * * * (or daily/weekly via the UI)
• Task: Run the curator skill

Results land in the Codex app's Triage view. Useful if you don't want to manage a local cron job, but only works against wikis the Codex Cloud runner can reach over the network.

Docs:

• Codex CLI reference
• Codex app Automations
• Non-interactive mode (codex exec)