Claude DuckDB Skill Guide (2026)

The prompt

Use the write-script-duckdb skill. I have ./orders.csv (~2M rows, columns: order_id, customer_id, country, amount, created_at). Write DuckDB SQL that returns the top 10 countries by total revenue, with order count and average order value, sorted by revenue descending. Use the FROM 'file.csv' shorthand - DuckDB auto-detects the schema. Don't create a table.

What slides.md looks like

-- DuckDB auto-detects types and headers from the CSV
SELECT
    country,
    count(*)            AS orders,
    sum(amount)         AS revenue,
    round(avg(amount),2) AS avg_order_value
FROM 'orders.csv'
GROUP BY country
ORDER BY revenue DESC
LIMIT 10;

-- Override detection only if needed:
-- FROM read_csv('orders.csv', header = true, sample_size = 20_000)

Pairs with

The prompt

Use the write-script-duckdb skill. Join ./events.csv (event_id, user_id, ts, action) against ./users.parquet (user_id, plan, signup_country) on user_id. Return action counts per plan tier for the last 30 days. Use read_csv() and read_parquet() in the same FROM clause. Files stay on disk; no CREATE TABLE.

What slides.md looks like

SELECT
    u.plan,
    e.action,
    count(*) AS n
FROM read_csv('events.csv') AS e
JOIN read_parquet('users.parquet') AS u
    ON e.user_id = u.user_id
WHERE e.ts >= now() - INTERVAL 30 DAY
GROUP BY u.plan, e.action
ORDER BY u.plan, n DESC;

Pairs with

The prompt

Use the write-script-duckdb skill. I have ./api_dump.json where each record has { id, customer: { name, tier }, items: [ { sku, qty } ] }. Write DuckDB SQL that returns one row per item with id, customer name, tier, sku, qty. Use read_json(), the -> / ->> JSON operators, and UNNEST for the items array.

What slides.md looks like

SELECT
    j.id,
    j.customer->>'name' AS customer_name,
    j.customer->>'tier' AS tier,
    item->>'sku'        AS sku,
    (item->>'qty')::INT AS qty
FROM read_json('api_dump.json') AS j,
     UNNEST(j.items) AS t(item);

-- For newline-delimited logs use read_ndjson('logs.ndjson')
-- or read_json(..., format = 'newline_delimited')

Pairs with

The prompt

Use the write-script-duckdb skill. Query this remote Parquet over HTTPS: https://example.com/data/nyc_taxi_2026.parquet. Return average trip distance and total fare by payment_type, for trips over $50. Read it straight from the URL with read_parquet() - the httpfs extension auto-loads on first remote read, so no manual INSTALL is needed.

What slides.md looks like

-- httpfs auto-loads on the first https:// read
SELECT
    payment_type,
    count(*)               AS trips,
    round(avg(trip_distance), 2) AS avg_miles,
    round(sum(fare_amount), 2)   AS total_fares
FROM read_parquet('https://example.com/data/nyc_taxi_2026.parquet')
WHERE fare_amount > 50
GROUP BY payment_type
ORDER BY total_fares DESC;

-- Glob a folder of remote files and keep the source:
-- SELECT *, filename FROM read_parquet('s3://bucket/y=2026/*.parquet')

Pairs with

The prompt

Use the write-script-duckdb skill. From ./daily_revenue.csv (date, region, revenue), compute per region: the 7-day moving average of revenue, the running cumulative total, and rank each day within its region by revenue. Use AVG/SUM OVER with frame clauses and row_number() OVER. Use QUALIFY to also return only each region's single best day.

What slides.md looks like

SELECT
    date, region, revenue,
    round(avg(revenue) OVER w7, 2) AS revenue_7d_avg,
    sum(revenue)       OVER wrun   AS revenue_running,
    row_number()       OVER wrank  AS day_rank
FROM 'daily_revenue.csv'
WINDOW
    w7   AS (PARTITION BY region ORDER BY date
             ROWS BETWEEN 6 PRECEDING AND CURRENT ROW),
    wrun AS (PARTITION BY region ORDER BY date),
    wrank AS (PARTITION BY region ORDER BY revenue DESC)
QUALIFY day_rank = 1;   -- best day per region only

Pairs with

The prompt

Use the write-script-duckdb skill. From ./sales.csv (country, year, population) build a cross-tab: one row per country, one column per year, cells = sum(population). Use DuckDB's PIVOT statement. Then show me the inverse with UNPIVOT to fold a wide year-columns table back into (country, year, population) long form.

What slides.md looks like

-- Wide cross-tab: years become columns
PIVOT (FROM 'sales.csv')
ON year
USING sum(population)
GROUP BY country;

-- Inverse: fold year columns back to long form
UNPIVOT wide_sales
ON "2024", "2025", "2026"
INTO
    NAME  year
    VALUE population;

Pairs with

The prompt

Use the write-script-duckdb skill. Read ./raw_events.csv, drop rows where user_id IS NULL, cast ts to TIMESTAMP, keep only the columns user_id, ts, action, and write the cleaned result to ./clean/events.parquet using COPY ... TO with FORMAT parquet. Then also show the PARTITION_BY variant that writes one folder per action.

What slides.md looks like

-- Single cleaned Parquet file
COPY (
    SELECT user_id, ts::TIMESTAMP AS ts, action
    FROM 'raw_events.csv'
    WHERE user_id IS NOT NULL
) TO 'clean/events.parquet' (FORMAT parquet);

-- Hive-partitioned: one sub-folder per action value
COPY (SELECT * FROM 'raw_events.csv')
TO 'clean/events' (FORMAT parquet, PARTITION_BY (action));

Pairs with

The prompt

Use the write-script-duckdb skill. I have ./app.sqlite with tables users(id, country) and sessions(user_id, started_at, duration_s). Attach it as a SQLite database and return median session duration and session count per country. Use INSTALL sqlite; LOAD sqlite; ATTACH ... (TYPE sqlite); then query the attached tables by qualified name.

What slides.md looks like

INSTALL sqlite;   -- once per environment
LOAD sqlite;
ATTACH 'app.sqlite' AS app (TYPE sqlite);

SELECT
    u.country,
    count(*)                        AS sessions,
    median(s.duration_s)            AS median_seconds
FROM app.sessions AS s
JOIN app.users    AS u ON u.id = s.user_id
GROUP BY u.country
ORDER BY sessions DESC;

Pairs with

The prompt

Use the write-script-duckdb skill. Create a typed DuckDB table 'transactions' (txn_id BIGINT, amount DECIMAL(12,2), currency VARCHAR, txn_date DATE) and bulk-load ./txns.csv into it with COPY ... FROM. The CSV has a header row and a non-default date format YYYY-MM-DD. After loading, return the row count and the sum by currency.

What slides.md looks like

CREATE TABLE transactions (
    txn_id   BIGINT,
    amount   DECIMAL(12,2),
    currency VARCHAR,
    txn_date DATE
);

COPY transactions FROM 'txns.csv'
    (HEADER true, DATEFORMAT '%Y-%m-%d');

SELECT currency, count(*) AS n, sum(amount) AS total
FROM transactions
GROUP BY currency
ORDER BY total DESC;

Pairs with

The prompt

Use the write-script-duckdb skill. From ./budget.xlsx, read the sheet named 'FY26' with the first row as headers, and return total budget and total actual by department, plus the variance. Use read_xlsx() with the sheet and header options - DuckDB infers the Excel extension automatically, no manual INSTALL.

What slides.md looks like

SELECT
    department,
    sum(budget)            AS budget,
    sum(actual)            AS actual,
    sum(actual) - sum(budget) AS variance
FROM read_xlsx('budget.xlsx', sheet = 'FY26', header = true)
GROUP BY department
ORDER BY variance;

-- .xlsx only; .xls is not supported by the reader.

Pairs with