Build an Analysis Dataset

For: Analytics Publishers

Use this page to create reusable dataset structures in an existing Analytics Project. All design, creation, and verification happens through your AI Agent and Culsma PAWS MCP; there is no Analytics Web/App editor in the current pilot.

Schema, table, and view creation are direct audited MCP write operations. The AI Agent must show the complete design and wait for explicit approval before executing them.

Before you start

You need:

• the exact Analytics Project selected and verified;
• an MCP connection with lab.analytics.write;
• an active Publisher grant for that Analytics Project; and
• a clear definition of the dataset’s purpose and one-row meaning.

If you will load rows after creating the structure, you also need a Writer grant. If you will query it yourself, you need Reader or Writer access.

Start in your AI Agent

Copy and adapt this request:

Design a reusable dataset

In Analytics Project [PROJECT NAME OR ID], propose a dataset for [PURPOSE]. One row represents [ROW MEANING]. Discover the existing schemas and tables first. Show the proposed schema, table, ordered columns, user-facing meaning, data type, required or optional status, and primary key. Explain how the design supports the intended questions. Do not create or change anything until I approve the complete design.

The AI Agent should:

1. state the selected environment, workspace, and Analytics Project ID;
2. confirm your active Publisher access;
3. discover existing schemas and tables to avoid collisions;
4. clarify the dataset’s row grain, identifiers, units, dates, nullability, and update pattern;
5. propose valid, stable object and column names; and
6. stop before any direct MCP write.

Review the dataset design

Design item	What to check
Analytics Project	Correct canonical project identifier.
Schema and table	Clear names that do not collide with existing objects.
Row meaning	One row has one unambiguous interpretation.
Columns	Name, meaning, unit, supported type, nullability, and order.
Identifier	Primary key or other uniqueness rule matches real data.
Dates and time	Date versus timestamp and time-zone meaning are explicit.
Future maintenance	Append, replacement, correction, and duplicate behavior are understood.

Culsma PAWS supports common analytics types such as text, integers, numeric values, booleans, dates, timestamps, UUIDs, and structured JSON. Ask the AI Agent to select the smallest type that preserves the meaning of the data.

Create the structure

Create dependent objects in a controlled sequence:

1. approve creation of the schema;
2. verify the schema appears in the project;
3. approve creation of one table with the exact reviewed columns;
4. fetch the table information and compare every column; and
5. only then continue to another table or view.

Approve one table

I approve creating schema [SCHEMA] and table [SCHEMA.TABLE] in Analytics Project [PROJECT ID] with exactly the ordered columns shown above. Create the schema first, verify it, then create only this table and return the audit and read-back evidence.

If an existing table is encountered, do not silently reuse it. The existing columns, types, nullability, order, and primary keys must match the approved design exactly before it is treated as the same table.

Create a reusable view

A view is useful when Readers repeatedly need the same filtered or joined result without modifying source data.

Ask the AI Agent to:

1. inspect the source tables and columns;
2. show the exact read-only SELECT or WITH definition;
3. dry-run the query as a read operation;
4. explain the row meaning and dependencies; and
5. stop before creating the view.

The view definition must remain read-only and may use only objects accessible to the Analytics Project Publisher role.

Review the evidence

For each created object, require:

• Analytics Project ID and environment;
• qualified schema, table, or view name;
• operation status and duration;
• ordered table columns, types, nullability, and primary-key flags;
• exact view query when a view is created;
• query or DDL audit ID and audit status; and
• a post-write schema or table listing that confirms the object exists.

This evidence is the only current confirmation surface; there is no Web/App dataset browser.

More requests to try

Inspect Analytics Project [PROJECT ID] and propose a table for one measurement per sample. Do not create it.

Compare my proposed columns with the current table [SCHEMA.TABLE] and report every difference in type, nullability, order, and primary key.

Propose a read-only view that returns the latest measurement for each sample. Show and dry-run the exact SQL before asking to create the view.

If dataset creation cannot continue

What you see	What to do
Publisher access is missing	Ask the Workspace Owner for a Publisher grant for the exact Analytics Project.
The schema name belongs to another project	Choose a project-owned schema; do not adopt or overwrite another project’s object.
The table already exists with different columns	Stop and decide whether to use a new name or perform a separately reviewed structural change.
A data type is unsupported	Choose a supported type that preserves meaning; do not silently coerce the data.
A primary key would reject real duplicates	Revisit the row grain and identifier before creation.
The view query is not read-only	Redesign it as a single SELECT or WITH query.
The write succeeds but audit finalization is pending	Keep the returned audit ID and verify the object through read-back before continuing.
Read-back differs from the approved design	Stop before loading data and report the exact object plus the mismatch.

You have completed the dataset structure when every created object matches the approved design and has MCP audit and read-back evidence.

To add rows, continue to Maintain Analytics Data. To query the dataset, continue to Explore and Analyze Data.