Move 2: the reasoning slice - shape, schema, sample, control
Lesson 3 of 6 · What the model can and cannot see
1 · Learn the move · Template pattern
You do not reason over the file. You reason over a slice, and a good slice has exactly four parts. Shape: rows, columns, dimensions, file count - the size and outline of the thing. Schema: column names, dtypes, units, factor levels - what each field means. Sample: the first and last rows, or one representative record - what the data actually looks like. Control: one number you already know, placed in the slice, so a wrong answer is visible immediately. A 40 GB alignment file becomes a flagstat table and a coverage summary; a 384-well run becomes a plate map and five summary rows. Handing over more does not help and usually hurts - the model reasons better over 60 honest rows than 60,000 truncated ones. And a slice with no control is how a scientist gets a confident wrong answer and quits. The template makes the slice a habit instead of a judgment call.
Before I ask any analysis question, here is the reasoning slice of my dataset. SHAPE: [rows x cols, or dimensions, file count]. SCHEMA: [each column: name, dtype, units, allowed values]. SAMPLE: [first 5 and last 5 rows, or one full record]. CONTROL: [one value I know independently, e.g. sample X reads Y within Z]. First, restate the dataset in your own words and state what the control lets you check. If the slice is missing something you need, ask - do not assume it.
2 · Your turn. You write the prompt
Your 384-well plate reader kinetics run produced an 80 MB export, and pasting chunks of it into a chat has produced three different wrong summaries in a row - each one confidently misreading truncated rows. Write the prompt that replaces the paste-and-pray with a four-part slice the model can actually reason over.
Remember: the AI sees only your prompt, not this page. If the situation isn't in your prompt, it doesn't exist.
Optional. These shape the output when you run your prompt below, not your score.