Dataset First Look

The first ten minutes with a new dataset decide whether the analysis is worth anything, and they are the minutes most often skipped. This skill runs a fixed first pass over whatever you can give it: a header and a few rows pasted into chat, a table copied from a spreadsheet, or a file attached to the conversation. One profile in one shape, so two datasets compare and a second look diffs against the first. It needs only what you paste or attach, so it runs the same on claude.ai, in Claude Code, and in Claude Cowork, where /first-look runs it over a file on disk and saves the result.

When to use this skill

• An extract arrived from another team and nobody can tell you what is in it.
• You are about to query an unfamiliar table and want the traps listed first.
• A number came out wrong and you suspect the source data rather than the logic.
• You need a written record of the dataset's condition to attach to an analysis.

Instructions

1. Say what you inspected, first and plainly. One line naming the input and its limits: "23 pasted rows out of a stated 1.4M", "the full attached CSV, 8,412 rows", "a header and 5 rows; no row count supplied". Every later statement inherits that scope. If no row count was given and you cannot see one, write "row count not supplied" and never estimate it.
2. Report the shape. Column count, row count if known, file or sheet name, and whether the header row looks like a header (names, not values) or like data that lost its header.
3. Build the column inventory as a table, one row per column: name, inferred type, evidence, example values, confidence. Infer from values, never from the name; a column called date holding 20240113 is an integer encoding a date, and both facts belong in the row.
4. Report null and blank patterns. Distinguish four things people conflate: a true null, an empty string, whitespace only, and a literal NULL, N/A, -, or none typed as text. Say whether nulls cluster (one date range, one segment) rather than scattering. Clustering is a finding; scatter usually is not.
5. Report cardinality. Per column, distinct values in what you can see: one value, a handful (a category), many with repeats (a dimension), or unique per row (a key or free text). Where the sample is partial, say "in the rows shown".
6. List suspicious columns. Go through this list every time, and say when a check found nothing: mixed types in one column; sentinel values standing in for missing data (-999, 0 in a price column, 1900-01-01); leading or trailing spaces; inconsistent categories (US, us, USA, United States); numbers stored as text; dates in more than one format; encoding damage; and IDs long enough to have lost precision in a spreadsheet.
7. Assess duplicate-key risk. Name the column or combination that looks like the intended grain, say whether it is unique in the rows you can see, and say plainly that uniqueness in a sample is not uniqueness in the table. Give the check that settles it.
8. Flag outliers worth a look, not every extreme value: impossible values (negative quantities, future timestamps), values orders of magnitude from their neighbors, and dates outside the range the dataset claims to cover. Say what each would do to a mean or a sum.
9. Write the questions to answer before analyzing. Five to ten, each answerable by a person or a query, ordered by how badly a wrong answer would break the analysis: what is one row, what period does this cover, are rows updated in place or only appended, what does a null here mean, what was filtered out before it reached you.
10. Close with a one-line verdict: ready to analyze, ready with named caveats, or blocked until a question is answered.

Output

One profile in the order above: scope line, shape, column inventory table, nulls, cardinality, suspicious columns, duplicate-key risk, outliers, questions, verdict. Save it beside the data, which is what /first-look does for you, and hand the questions to whoever owns the source. When the data is a spreadsheet model rather than a flat extract, the spreadsheet-formula-auditor skill audits the formulas behind the numbers; once it is clean, chart-chooser picks how to show it.

Example

Excerpt from a profile of a pasted order extract:

Inspected: 30 pasted rows of a stated 240,000; 11 columns. Statements below cover the 30 rows shown.
 
| Column | Inferred type | Evidence | Example | Confidence |
| --- | --- | --- | --- | --- |
| order_id | integer key | unique across 30 rows, monotonic | 100482 | high |
| country | category, dirty | 4 spellings of 2 countries | "us", "USA" | high |
| discount | number as text | separators, one "N/A" | "1,250" | high |
| shipped_at | date, sentinel | 6 rows read 1900-01-01 | 2026-08-14 | medium |
 
Suspicious: `shipped_at` uses 1900-01-01 where a spreadsheet wrote an empty date. Treat as unshipped.
Duplicate-key risk: `order_id` is unique in 30 rows. Confirm on the full table before joining on it.
 
## Questions before analyzing
1. Is one row an order or an order line? `discount` looks per-line.
2. Are orders updated in place after shipping, or appended as new rows?

Where this sits in the analyst set is covered in Claude skills for data analysts; the analysis-reviewer agent checks the finished analysis that this profile starts.