Daniel StoyanovPrismClaude Agent SDK · Python

Claude Agent SDK · Python · DuckDB · read-only

Prism

Plain English in, verified SQL and a chart out

A data analyst agent that turns a question into SQL, runs it in a sandbox, checks the result against the question it was asked, and renders a chart. If the query returns something implausible it re-queries instead of charting nonsense.

Take a look inside3 scenarios inside

What makes it special

read → run → check → chart
  1. 01Read

    Reads the schema before it writes SQL

    Prism is told never to guess a column name, so it starts by pulling tables, types, row counts and the notes a type cannot carry, like refunds living on the order rather than the line. If a query still hits a missing column, the error comes back with a hint to re-read the schema, not to try another guess.

    describe_schema()agents/prism-analyst/prism/warehouse.py

  2. 02Run

    Read-only, three layers deep

    A parser refuses anything but a single SELECT before DuckDB sees it, and the connection itself is opened read-only in case that parser has a bug. The agent has no shell, file or web tools at all, so the only way to the data is one SELECT at a time.

    run_sql()agents/prism-analyst/prism/warehouse.pyagents/prism-analyst/prism/analyst.py

  3. 03Check

    Checks the answer against the question

    Every result comes back with computed warnings: a group key that never reaches the aggregate, a rate landing exactly on ½, a denominator under 25. That is how it catches a tidy 50% retention rate that is really 4,312 orders divided by 2,156 customers.

    inspect_result()agents/prism-analyst/prism/selfcheck.py

  4. 04Chart

    Won’t chart a number it doesn’t trust

    It is instructed to say so and re-query rather than plot a result it distrusts. The chart tool also refuses forms that mislead: a line across categories with no order, a donut with negative or more than six slices, a fourth series the validated palette has no colour for.

    render_chart()agents/prism-analyst/prism/charts.py

Inside the demo

3 scenarios
  1. 01

    Revenue by channel, Q2

    Schema first, one readable query, then the chart the data shape asks for.

    question inrevenue by channel for Q2

  2. 02

    Self-check kills its own query

    The first query returns a clean-looking 50% - and cannot answer the question. Prism says so and re-queries.

    question inwhat is our 90-day retention for customers who frist ordered in Q1, split by channel?

  3. 03

    Query fails on a column that does not exist

    run_sql comes back ok:false with a binder error. Prism re-reads the schema instead of guessing a column name.

    question intop 5 products by refund rate last month

Under the hood

python

Architecture

stack
Claude Agent SDK · Python
pattern
Agent loop with sandboxed code execution
why this, and what it costs
Letting the model write and run code beats hand-built query tools the moment questions get compositional: you cannot pre-build a tool for every aggregation. The price is a sandbox you must actually isolate, and output you must validate before trusting.

Tools

3
  • describe_schemasafe

    Return tables, columns and types available to query.

  • run_sqlsafe

    Execute a read-only SQL query in the sandbox and return rows.

  • render_chartsafe

    Render a chart from a result set.

Source

runnable
real source
agents/prism-analyst/prism/analyst.py
packages
  • claude-agent-sdk
  • duckdb
  • pandas
written against
api_documentation/claude-agent-sdk-python.md

Watch it throw out its own 50%.

Inside, each question runs step by step: Prism reads the schema, each query lands with its rows and timing, and a result it rejects stays on screen, struck through, while the chart is drawn from the one that replaced it.

Take a look inside