LLM Classification Is Feature Engineering

Who Should Read

A backend developer running LLM prompt classifiers who can't get calibrated probabilities or threshold control.

Core Mechanics

• LLM classifier probabilities can't be trusted
  • LLM verdicts are usually hard labels like Ironic/Not.
  • Token logprobs and self-reported confidence have no reason to be calibrated.
  • So trading off precision and recall in a principled way is hard
• No guarantee the LLM uses your structured data
  • Pasting structured data into the prompt doesn't mean the LLM uses it.
  • The LLM also doesn't know your population's base rate of positives
• Put a logistic regression on top of the verdict
  • Treat the LLM verdict as one feature and fit coefficients on training data.
  • With a binary verdict, the fitted probability is just the empirical positive rate per verdict
• Real probabilities give you threshold control
  • Other covariates can simply be added to the regression.
  • The model can also be swapped for xgboost or a neural network

Community Reactions

• Suggestion to automate prompt generation
  • A commenter proposed a strong LLM writing prompts and a weak LLM classifying, feeding a small ML model.
  • They would regenerate prompts from the most misclassified examples and add them as features, repeating.
• Experience using LLM decisions as features
  • A commenter doing LLM-as-a-judge work said feeding LLM decisions into a classic ML model for the final call works really well.
• Objection that a megaprompt baseline is missing
  • A commenter said the post should compare against one prompt listing the properties and asking for reasoning then an answer.
  • They expect it to get close to the logistic classifier and generalize better out of distribution.
• Tip to prefer binary questions
  • One commenter found binary questions beat multi-class, and answer consistency became an extra pipeline check.
• Point that logprobs may not reflect intent
  • A commenter noted structured output may mask disallowed tokens so probabilities differ from the model's original weights.

How to Apply

• Add a logistic regression to your current LLM classifier
  • Run the LLM on a labeled validation set and compute the positive rate per verdict.
  • Use those rates as probabilities and compare the Brier score to before
• Turn misclassifications into sub-question features
  • Read 10 wrong cases and turn shared traits into yes/no questions.
  • Extract them all as bool fields in one structured output call
• Merge rule-based features into the same regression
  • Add computable values like length, URL presence, and exclamation count next to the LLM features.
  • Compare F1 for verdict only, +LLM features, and +rule features
• Set the threshold from business cost
  • Once you have probabilities, plot the precision-recall curve.
  • Pick the threshold that matches your false-positive cost

Try It

# Paste into Claude or ChatGPT (simplified version of the article's method: feature design only, no code)

You are a classification feature designer. For the task below, write yes/no feature questions an LLM can extract in one call.

Task: {{classification task description}}

Cases the LLM currently gets wrong:
{{5-10 misclassified cases with correct labels}}

Instructions:
1. Summarize common traits of the misclassified cases in 3 lines.
2. Write 10 bool feature questions answerable independently of the verdict.
3. Write a pydantic Field description for each question.
4. Check whether the questions overlap.

Next step: run the questions on your validation data and compare F1 and Brier against the verdict-only model.

Terminology

Related Papers

• The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge

  Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.

• 90% of US working hours are spent on tasks not covered well by AI benchmarks

  Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage

• Postgres 19 delayed after AI reviewers find more bugs than humans can fix

  LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted

• Show HN: Gemini 4 Argon plays StarCraft

  LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.

• A Plausible Explanation Is Enough to Mislead Multimodal Judges

  Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.

• Neural networks are clean

  Looking at neurons directly shows most have one main job, challenging superposition.

Related Resources

• https://minimallysufficient.com/posts/llm-classification-is-feature-extraction/
• https://softwaredoug.com/blog/2025/01/21/llm-judge-decision-