profile-data

name	profile-data
description	Profile CSV files in data/ for schema, quality, null counts, unique counts, sample values, and join-key candidates
tags	["data-quality","profiling","exploration"]
version	1.0.0

Profile Data

Use this workflow to understand the schema and quality of data files BEFORE building transforms against them. It's the first step when onboarding a new project or when the user drops a CSV into data/ and wants to know what's inside.

Tool you will call

• profile_data — profiles CSV files in data/

When to use

Trigger when the user says things like:

• "what's in data/"
• "profile the customers file"
• "show me the schema"
• "what's the join key"
• "how many nulls in X"

Also use PROACTIVELY at the start of a session when:

1. The user is new to the project and asks an analytical question
2. The user mentions a table you haven't seen yet
3. You're about to build a new transform and need to verify column types

Phase 1 — Decide scope

The tool has two modes:

Overview mode (file_path="")

Lists ALL CSVs in data/ with:

• Row count per file
• Column name + type summary per file
• A "potential join keys" section showing columns that appear in 2+ files, with type mismatch detection flagged with TYPE MISMATCH.

Use this when you don't know what files exist or when the user asks an open-ended "what's available" question.

Detail mode (file_path="data/<name>.csv")

Detailed profile of ONE CSV:

• Row count
• Per-column: type, null count / total, unique count, up to 5 sample values

Use this when the user mentioned a specific file OR when you need to verify the shape of a table before building a transform against it.

Phase 2 — Call the tool

Args:

• file_path (optional): path to a specific CSV (e.g. "data/customers.csv"). Leave empty for overview mode.

Phase 3 — Interpret the results

Join keys with type mismatches

When the overview output flags TYPE MISMATCH, DO NOT silently ignore it. This is the most common cause of broken JOINs: one file has customer_id as VARCHAR, another as BIGINT, and the JOIN silently returns zero rows.

Tell the user about the mismatch and suggest either:

1. Adding a CAST in the transform's SQL, OR
2. Normalizing the CSV upstream so both files agree on the type.

Null counts

If a column has high null density, flag it before the user builds a transform that assumes it's populated.

Sample values

Use the sample values to verify your mental model of the column (e.g., order_status has values Pending, Shipped, Cancelled — not a boolean).

Phase 4 — Next steps

After profiling, suggest next actions based on what you found:

• Clean schema → use bootstrap-semantic-model skill to auto-generate a semantic layer
• Complex shape / many files → use search_project_files to find existing transforms that already work with this data
• Obvious join keys → use build-pipeline-node skill to draft a new transform that JOINs them

Output location

profile_data is read-only — it writes nothing to disk. All results are returned as the tool response.