find-events - SKILL.md Agent Skill

name: find-events description: "This skill should be used when the user asks 'What events are happening in Berlin this week?', 'Find me art exhibitions and food festivals in Berlin next weekend', or wants to discover upcoming Berlin art and food events. Ingests event sources into qurl, runs semantic search, checks Google Calendar for conflicts, and produces a curated, relevance-ranked list of events with location context." argument-hint: "[days ahead, e.g. '7' or 'this weekend']" allowed-tools: ["Read", "Bash", "WebSearch", "WebFetch", "Grep", "Glob", "Agent", "mcpclaude-in-chromenavigate", "mcpclaude-in-chromeread_page", "mcpclaude-in-chromeget_page_text", "mcpclaude-in-chrometabs_create_mcp", "mcpclaude-in-chrometabs_context_mcp"]

Find Berlin Events

Ingest Berlin event sources into qurl, run semantic search, check Google Calendar for conflicts, and produce a curated, relevance-ranked list.

Workflow

Step 1: Load User Settings

Read settings from .claude/berlin-events.local.md in the current project root — the single canonical location, the same path the /berlin-events:init command writes to. Do not look anywhere else (no ~/.claude fallback); a second location only creates ambiguity about which file wins. Extract:

neighborhood: User's Berlin neighborhood (for travel context)
interests: Art, food, or both (default: both)
calendar_id: Google Calendar ID (default: primary)
lookahead_days: How many days ahead to search (default: 14)

If the file is absent, assume defaults (neighborhood=Mitte, interests=art+food, calendar=primary, lookahead=14) and tell the user they can run /berlin-events:init to create .claude/berlin-events.local.md in this project.

Step 2: Determine Date Range

Parse the optional argument for date range:

No argument: today through 14 days ahead
Number (e.g., "7"): today through N days ahead
"this weekend": upcoming Saturday and Sunday
"next week": Monday through Sunday of next week

Calculate exact dates using today's date. Only include future events.

Step 3: Ingest Sources into qurl

Scrape each priority source and ingest it into the local qurl database.

Do not background the qurl add calls. qurl writes to a single sqlite file; concurrent writers hit database is locked and silently drop sources (seen in real runs: co-berlin dropped). Run the ingest calls sequentially — scraping is the slow part, the writes are fast. The bun run … | qurl add pipe below is already sequential; keep it that way.

PLUGIN_ROOT="${CLAUDE_PLUGIN_ROOT:-plugins/berlin-events}"

ingest() {
  local url="$1" tags="$2"
  bun run "$PLUGIN_ROOT/scripts/extract-content.js" "$url" \
    | qurl add "$url" --source berlin-events --tags "$tags"
}

# Art sources
ingest "https://www.indexberlin.com/events/list/"                                         "art"
ingest "https://www.kw-berlin.de/en/events"                                               "art"
ingest "https://berlinischegalerie.de/programme/kalender/"                                "art"
ingest "https://www.artatberlin.com/en/calendar-for-vernissagen-exhibitions-events/"      "art"
ingest "https://co-berlin.org/de/programm/kalender"                                       "art"
ingest "https://kunstleben-berlin.de/events/"                                             "art"

# Food / broad sources
ingest "https://www.berlin.de/en/events/"                  "food"
ingest "https://www.visitberlin.de/en/event-calendar-berlin" "art,food"

qurl deduplicates by URL + content hash — re-running is safe.

Step 4: Embed

Generate vector embeddings for any newly ingested documents:

qurl embed

Step 5: Search with qurl query

Tool note — use qurl, NOT qmd. This pipeline searches the qurl event-scrape database. qmd is a different engine over your markdown notes corpus — it does not contain any scraped events and will return nothing relevant. Never substitute a qmd command for a qurl one, even where they share a subcommand name.

Run a short (3–5 word) BM25 query first; fall back to vsearch if fewer than 5 hits. For the full command comparison, --source/--tag flag behaviour, vsearch grep pipeline, and relevance-filter token derivation: → references/qurl-search.md

qurl search "Berlin art exhibition opening" --source berlin-events --limit 20

Relevance-filter the results against date/event keywords. If fewer than 5 relevant hits, proceed to Step 5b.

Step 5b: Web Search Fallback (only if query < 5 results)

"Berlin art events this week [date range]"
"Berlin food events this week [date range]"
"Berlin exhibition openings [date range]"

Step 6: Check Google Calendar

Check for scheduling conflicts over the lookahead window.

Interactive sessions: prefer the Google Calendar MCP tools if available (mcp__claude_ai_Google_Calendar__list_events for the date range). No install or OAuth needed.
Headless / no MCP (the weekly launchd job): fall back to gogcli (binary gog):
```
gog calendar events --from today --days 14 --json
```
If gogcli is not installed or not authenticated, see references/calendar-add.md.

Parse output to identify busy time slots. Flag events that overlap with existing entries.

Step 7: Add Location Context

For each event, note the neighborhood (e.g., "Kreuzberg", "Mitte", "Charlottenburg") and compare with the user's neighborhood:

Same neighborhood: "Near you"
Adjacent: "~15 min by transit"
Far: "~30+ min by transit"

Use general Berlin geography knowledge. Do not call external routing APIs.

Step 7.5: Drop already-shown events (dedup) and load taste feedback

Initialize the state DB once (idempotent), then filter out events surfaced in previous runs so the user never sees the same suggestion twice. Resolve $BUN first:

BUN=$(command -v bun 2>/dev/null || echo "$HOME/.bun/bin/bun")
PLUGIN_ROOT="${CLAUDE_PLUGIN_ROOT:-plugins/berlin-events}"
$BUN run "$PLUGIN_ROOT/scripts/events-db.ts" init

Write your curated candidate events to a temp JSON array (each object needs title, date, venue, plus any other fields you carry), then filter:

echo "$CANDIDATES_JSON" | $BUN run "$PLUGIN_ROOT/scripts/events-db.ts" filter

Only the returned (unseen) events proceed. Also load recent taste feedback to inform ranking:

$BUN run "$PLUGIN_ROOT/scripts/events-db.ts" recent-feedback --limit 20

Bias ranking toward venues/categories the user marked went and away from those marked skip.

Step 7.6: Weather Scoring

Run the weather script for the date range. Pass --config if the settings file has a weather block; omit it otherwise.

BUN=$(command -v bun 2>/dev/null || echo "$HOME/.bun/bin/bun")
PLUGIN_ROOT="${CLAUDE_PLUGIN_ROOT:-plugins/berlin-events}"
WEATHER_CFG_JSON=$($BUN run "$PLUGIN_ROOT/scripts/read-weather-config.ts" .claude/berlin-events.local.md 2>/dev/null || echo "")
if [ -n "$WEATHER_CFG_JSON" ]; then
  WEATHER_JSON=$($BUN run "$PLUGIN_ROOT/scripts/weather-gate.ts" --from "$DATE_FROM" --to "$DATE_TO" --config "$WEATHER_CFG_JSON")
else
  WEATHER_JSON=$($BUN run "$PLUGIN_ROOT/scripts/weather-gate.ts" --from "$DATE_FROM" --to "$DATE_TO")
fi

DATE_FROM and DATE_TO are the ISO-8601 start/end dates calculated in Step 2.

Classify each event as indoor/outdoor by keyword-matching venue, name, and description; apply outdoor_delta/indoor_delta per day; honour drop_outdoor as a hard removal. For full keyword lists, scoring algorithm, lake-suggestion rule, and weather-note header format: → references/weather-scoring.md

Step 8: Rank and Curate

Score events by:

Relevance to user interests (art/food)
Time fit — no calendar conflicts, within lookahead window
Proximity to user's neighborhood
Uniqueness — special/one-time events ranked higher than recurring
Source quality — primary sources and editorial picks ranked higher
Taste — prior went/skip feedback (Step 7.5)
Weather — apply outdoor_delta / indoor_delta from Step 7.6 (already done inline in 7.6; Step 8 uses the adjusted scores)

Step 9: Present Results

Include a summary at the top: "Found X events (Y art, Z food) for [date range]. N conflicts with your calendar."

Output events grouped by date. For the Google Calendar URL format and gogcli calendar create command: → references/calendar-add.md

## [Day, Date]

### [Event Name]
- **What**: [Brief description]
- **Where**: [Venue, Neighborhood] — [travel context from your location]
- **When**: [Time]
- **Category**: Art | Food
- **Link**: [URL]
- **Calendar conflict**: None | "Conflicts with [existing event] at [time]"
- **Add to calendar**: [Google Calendar link or gogcli command — see references/calendar-add.md]

---

After presenting, record the events you showed so they are not repeated next run:

echo "$PRESENTED_JSON" | $BUN run "$PLUGIN_ROOT/scripts/events-db.ts" record

To capture taste over time, the user can later mark an event with the /berlin-events:feedback command (resolves the event by title and records the verdict):

$BUN run "$PLUGIN_ROOT/scripts/events-db.ts" feedback --hash <event-hash> --verdict went|skip --notes "..."

Tips

If a source fails extraction, skip it and continue — do not abort
Prefer English-language sources but include notable German-only events
For food events: markets, pop-ups, food festivals, tastings, food-related workshops
For art events: openings, exhibitions, gallery walks, art talks, performances
Limit output to ~15–20 best events
See event-sources skill for validated source list and ingestion notes