NYC yellow taxi trips, January 2026

Contents

• What the dashboard shows
• End-to-end workflow
• Commands used
• Claude API usage metrics
• Conclusions

What the dashboard shows

Section	Chart type	Key insight
Takeaway sentence	Text with live-filled figures	3.5M trips, $104.4M charged, 62.3% paid by card, 6.3% from airports; busiest day Fri Jan 30 (140,853 trips)
Headline KPIs	Six number tiles	Avg fare $21.08 · avg distance 3.49 mi · median trip 13.5 min · card tip rate 21.2%
Trips per day	Bar chart, weekends muted, hover tooltips	Weekly rhythm and one sharp late-January dip
When people ride	Weekday × hour heatmap	Evening peaks Thu–Sat; late-night demand on weekends
Trip length	Horizontal bars by distance band	Most trips are 1–5 miles; 20+ mi trips average $112.19
How riders pay	Table	Card riders tip $4.11 on average; cash tips aren’t recorded

Data cleaning rules: the dashboard keeps only pickups in January 2026 and drops records with a zero or negative total, a distance of zero or 100+ miles, or a drop-off before the pickup. That left 3,517,748 trips.

End-to-end workflow

1. Profile the Parquet file with DuckDB

DuckDB read the Parquet schema (20 columns: pickup/drop-off timestamps, distance, fare, tip, payment type, airport fee, the new cbd_congestion_fee) and the date range. A few pickups fell outside January (from 2025-12-31 to 2026-02-01) and were filtered out.

Where does the data come from? The NYC Taxi & Limousine Commission (TLC) yellow taxi trip records, file yellow_tripdata_2026-01.parquet.

2. Aggregate 3.7M rows into five small CSVs

A dashboard can’t attach a 64 MB file, so DuckDB SQL reduced the data to five small summary tables (216 rows in total):

File	Rows	Contents
summary.csv	1	Month totals and rates
daily.csv	31	Trips, revenue, avg fare, avg miles per day
hourly.csv	168	Trips per ISO weekday × hour
payment.csv	5	Trips, avg total, avg tip per payment type
distance.csv	6	Trips and avg total per distance band

3. Create the Claude Dashboard

• Created a new artifact from the Dashboard artifact type on claude.ai.
• Uploaded each CSV to the artifact’s asset store.
• Wrote the dashboard’s documents in one atomic batch: the dash/meta title, five datasets/* source definitions, and the page files/index.html.
• Built the page in plain HTML with d3 v7. Each figure is tagged with data-source / data-field / data-where, so readers can click any number and see which CSV row it came from.

4. Visual check and fixes

A headless Chrome screenshot found two layout bugs, and both were fixed in the embed and in the live dashboard:

• The trip-length bars weren’t on one scale because the value column changed width from row to row. The value column now has a fixed width.
• The bottom card row left an empty third grid column. The grid is now an explicit two-column layout that drops to one column below 760px.

Commands used

Inspect the Parquet file

duckdb -c "DESCRIBE SELECT * FROM 'yellow_tripdata_2026-01.parquet';
SELECT count(*), min(tpep_pickup_datetime), max(tpep_pickup_datetime)
FROM 'yellow_tripdata_2026-01.parquet';"

Clean and aggregate to CSV (DuckDB SQL)

CREATE VIEW t AS SELECT * FROM 'yellow_tripdata_2026-01.parquet'
 WHERE tpep_pickup_datetime >= '2026-01-01' AND tpep_pickup_datetime < '2026-02-01'
   AND total_amount > 0 AND trip_distance > 0 AND trip_distance < 100
   AND tpep_dropoff_datetime > tpep_pickup_datetime;

COPY (SELECT count(*) trips, round(sum(total_amount),2) revenue, round(avg(fare_amount),2) avg_fare,
  round(avg(trip_distance),2) avg_miles,
  round(median(date_diff('second',tpep_pickup_datetime,tpep_dropoff_datetime))/60.0,1) median_minutes,
  round(sum(tip_amount) FILTER (WHERE payment_type=1)/sum(fare_amount) FILTER (WHERE payment_type=1),4) card_tip_rate,
  round(count(*) FILTER (WHERE payment_type=1)::double/count(*),4) card_share,
  round(count(*) FILTER (WHERE Airport_fee>0)::double/count(*),4) airport_share
  FROM t) TO 'summary.csv';

COPY (SELECT strftime(tpep_pickup_datetime,'%Y-%m-%d') AS "day", count(*) trips,
  round(sum(total_amount),2) revenue, round(avg(fare_amount),2) avg_fare,
  round(avg(trip_distance),2) avg_miles FROM t GROUP BY 1 ORDER BY 1) TO 'daily.csv';

COPY (SELECT isodow(tpep_pickup_datetime) dow, hour(tpep_pickup_datetime) AS "hour", count(*) trips
  FROM t GROUP BY 1,2 ORDER BY 1,2) TO 'hourly.csv';

COPY (SELECT CASE payment_type WHEN 0 THEN 'Flex fare' WHEN 1 THEN 'Credit card' WHEN 2 THEN 'Cash'
  WHEN 3 THEN 'No charge' WHEN 4 THEN 'Dispute' ELSE 'Other' END payment,
  count(*) trips, round(avg(total_amount),2) avg_total, round(avg(tip_amount),2) avg_tip
  FROM t GROUP BY 1 ORDER BY trips DESC) TO 'payment.csv';

COPY (SELECT CASE WHEN trip_distance<1 THEN 'Under 1 mi' WHEN trip_distance<2 THEN '1–2 mi'
  WHEN trip_distance<5 THEN '2–5 mi' WHEN trip_distance<10 THEN '5–10 mi'
  WHEN trip_distance<20 THEN '10–20 mi' ELSE '20+ mi' END band,
  min(trip_distance) lo, count(*) trips, round(avg(total_amount),2) avg_total
  FROM t GROUP BY 1 ORDER BY lo) TO 'distance.csv';

Claude Code tool calls (Artifact / ArtifactData)

Step	Tool call
Pick the artifact type	Artifact action: quickstart, intent: other
Create the dashboard	Artifact publish with the Dashboard type_url, title, auto_open: after_first_write
Read the type’s docs	Artifact action: read, path: artifact-type/reference/design-defaults.md (later exporting.md)
Upload data	Artifact publish with asset: true, one call per CSV, which returns /_blob/<id> URLs
Write the dashboard	ArtifactData action: batch (7 writes: dash/meta, 5 × datasets/*, files/index.html)
Fix the layout	ArtifactData action: update on files/index.html with if_version: 1

Claude API usage metrics

Usage summary (/usage equivalent)

Metric	Value
Model	claude-opus-5-5 (Opus 5.5), every request
Wall-clock time	16:14:43 → 16:25:12 UTC (≈ 10.5 min)
API requests	22
Input tokens (uncached)	44
Cache-write tokens	88,966
Cache-read tokens	1,786,426
Output tokens (incl. thinking)	29,567
Cache hit rate on input	≈ 95% (1.79M read vs 89K written)

Tool calls

Tool	Calls
Bash	10
Artifact	9
AskUserQuestion	4
ArtifactData	2
ToolSearch	1
Write	1
Read	1

Context window

Metric	Value
Context in use at the last request	≈ 112.9K tokens (112,095 cached + 825 new + 2 uncached)
Largest contributors	System prompt and tool schemas, the Dashboard type’s instructions and design references, the page HTML (~17 KB), and one rendered screenshot
Compaction	None needed

Conclusions

Claude dashboards can turn into a very powerful tool for fast creation of dashboards and data analysis.

However, it should not be unattended - the prompter must know what is the general view of the data being analyzed. Claude dashboards can turn into a very powerful tool

It was also very positive to see that there is traceability of what Claude did for analysing the data - useful for reproducibility