A Sketch on a Napkin, Cut Into Cloth: LLM-Generated Features from Unstructured Data

Opening Scene

A client can’t describe what they want in tailoring terms — they sketch a rough collar shape on a napkin, mention a jacket they saw in a film, gesture at how they want the shoulders to sit. None of that is a measurement or a pattern. It takes real skill to turn that loose, unstructured description into something with an actual cut and a seam allowance the workroom can use.

In Plain English

Large language models can now read genuinely unstructured data — free-text reviews, support tickets, contracts, doctor’s notes — and extract structured, model-ready features from it: a sentiment score, a list of mentioned product defects, a risk category, a summarized complaint type. What used to require painstaking manual tagging or was simply left unused can now be turned into usable features directly from raw text.

The Old Way

Unstructured text sat in the warehouse largely unused as a feature source, because turning it into structured signal required expensive, slow manual labeling.
Simple keyword-matching or rule-based rules extracted only crude signals from text, missing nuance, tone, and context an experienced human would easily catch.
Building a proper text-classification model just to extract one feature was often too costly to justify for anything but the highest-value use cases.

What’s Changing (and Why AI Is the Reason)

Large language models can now extract structured features directly from raw text through a plain-language prompt, functioning like a skilled tailor reading a client’s napkin sketch and vague description and turning it directly into a cutting instruction.
This dramatically lowers the cost of turning previously unused unstructured data — reviews, transcripts, notes — into genuinely useful features, opening up entire raw data sources that were, in practice, sitting unused on the shelf.
Because LLM outputs carry their own uncertainty and can be inconsistent, teams are building explicit review and validation steps around LLM-generated features, the same disciplined quality check from Article 16 applied to this genuinely new feature source.

The Metaphor, Fully Extended

Tailoring Element	LLM-Generated Feature Concept
A client’s rough napkin sketch and vague spoken description	Raw unstructured text — reviews, notes, transcripts
A skilled tailor interpreting that sketch into an actual cutting instruction	An LLM extracting a structured feature from unstructured text via a prompt
Fabric that sat unused because no one could interpret the request	Unstructured data that historically sat unused for lack of an economical way to extract signal
Checking the tailor’s interpretation against what the client actually meant	Reviewing and validating LLM-generated features before trusting them
A cutting pattern derived from conversation rather than a formal measurement	A structured feature — sentiment, category, risk score — derived from free text

For Beginners: What to Actually Do

Treat an LLM-generated feature as a draft interpretation, not a guaranteed-correct measurement — validate it before relying on it heavily.
Write clear, specific prompts when extracting features from text, since vague prompts produce vague, inconsistent outputs.
Spot-check LLM-extracted features against the original text periodically to catch drift or misinterpretation.
Recognize which unstructured data sources in your organization have likely been sitting unused and worth revisiting now.

For Practitioners and Leaders: The Deeper Layer

Invest in LLM-based feature extraction for unstructured data sources that were previously too costly to featurize manually.
Require validation and quality checks on LLM-generated features, since inconsistency and hallucination are genuine risks, not edge cases.
Track the cost and latency of LLM-based feature extraction at scale, since it behaves differently from traditional deterministic transformations.
Treat this as expanding the feature toolkit, not replacing the point-in-time and quality discipline the rest of this series has covered.

Quick Recap

Large language models can now extract structured, model-ready features directly from previously unstructured raw text.
Text data historically sat unused as a feature source because manual extraction was too slow and costly.
LLM-based extraction dramatically lowers that cost, but needs its own review and validation discipline.
A skilled tailor turning a napkin sketch into a real pattern is a fair picture of what an LLM now does with a paragraph of raw text.

Where This Fits in the Series

This article covered a genuinely new feature source. Article 18 turns to a more practical concern that applies to every feature covered so far — what all this cutting, fitting, and re-fitting actually costs.