interview-question-importer

name: interview-question-importer description: Import interview questions into interview/questions.md for the deep-learning-playbook repository. Use this skill when the user provides a screenshot, OCR result, note, or plain text containing interview questions and wants them classified into the existing sections of interview/questions.md, merged with deduplication, optionally folded into nearby existing questions as suffix expansions, and reported as added, skipped, expanded, or unmatched.

Interview Question Importer

Overview

Turn raw interview questions from an image or text block into clean updates to interview/questions.md.

The repository is a study notebook, so preserve the existing outline and keep edits pragmatic:

• classify into the most suitable existing section
• skip exact duplicates
• merge close variants into an existing question when that reads better than adding a new line
• report anything that does not fit an existing section

Workflow

1. Extract all questions from the input

Always start by reading the full user input carefully.

If the input is an image:

• read every visible question from the image directly
• fix only obvious OCR-style issues such as spacing, casing, or punctuation
• preserve technical terms such as ReLU, sigmoid, Cross Entropy, DIN, pooling

If the input is text:

• split it into individual questions
• keep the user's wording unless a very small cleanup makes the meaning clearer

Extraction rules:

• do not silently drop a line that looks like a question
• if one line contains one core question plus a follow-up angle, keep it as one question unless the prompts are clearly independent
• if an image is blurry or a phrase is ambiguous, make the smallest reasonable correction and mention the uncertainty in the final report

2. Read the repository taxonomy first

Open interview/questions.md before editing.

Find the most specific existing placement for each extracted question:

• prefer an existing subsection over a broad top-level section
• match by topic, not by superficial keywords alone
• reuse nearby clusters when several adjacent questions already cover the same concept family

Do not create a new section by default. If no suitable existing section is found, do not force the question into a weak match. Report it to the user as unmatched.

3. Choose exactly one action per question

For each extracted question, choose one of these outcomes:

skip

Use skip when an existing question is already effectively the same.

Treat as duplicate when:

• the wording is identical
• the wording differs slightly but the interview intent is clearly the same

Do not add a second copy just because the wording is shorter or longer.

expand

Use expand when the new question is very close to an existing one but adds a useful new angle.

Typical cases:

• existing question asks for the core concept, new question adds a comparison angle
• existing question asks for the method, new question adds a scenario or constraint
• existing question is broad, new question contributes a natural follow-up scope

When expanding:

• edit the existing question line directly
• keep the sentence natural and compact
• add the new angle as a suffix or rewrite the sentence into a slightly broader question
• do not turn one line into a paragraph

Example pattern:

• existing: 如何解决过拟合问题？
• new angle: 怎么在模型层面解决
• updated: 如何解决过拟合问题？有哪些数据、训练和模型层面的常见方法？

add

Use add when the question is genuinely new for that subsection.

Insertion rules:

• default to appending at the end of the target subsection
• if there is a clearly related local cluster, insert beside that cluster instead
• preserve the repository's numbered-list style
• renumber only within the affected subsection when needed

4. Keep question wording clean and interview-oriented

The goal is a question bank, not polished prose.

Prefer wording that is:

• concise
• technically precise
• easy to scan
• faithful to the original interview intent

Editing guidance:

• keep Chinese as the default language unless the surrounding section clearly uses English terms
• retain standard English technical identifiers where natural
• normalize obvious mistakes like sigmod to sigmoid only when the intent is unambiguous
• avoid adding answer content into interview/questions.md

5. Respect repository boundaries

• edit interview/questions.md only
• do not touch interview/answer/ unless the user explicitly asks
• do not reorganize headings unless the user requests structural cleanup
• do not modify unrelated files

Matching Heuristics

Use semantic fit over literal overlap.

Examples:

• loss functions, entropy, KL, softmax gradients -> DL 基础 / Loss 相关
• activations, ReLU vs sigmoid, dropout, batch norm, overfitting mechanisms -> DL 基础 / DL 网络机制
• LLM sampling, quantization, fine-tuning, deployment -> matching LLM 类通用 subsections
• RAG retrieval, rerank, chunking -> Agent / RAG 应用 / RAG

If a question spans multiple sections, prefer the section where a reviewer would most naturally look for it later.

Final Report

After editing, always report the outcome in four buckets:

1. 实际新增
2. 追加到已有问题
3. 跳过
4. 未找到合适分区

For each item, include:

• the normalized question text
• the target section or matched existing question when applicable

If any extraction was uncertain, add one short note describing the assumption.

Default Operating Mode

When this skill triggers, default to executing the edit directly instead of only describing a plan.

Use this order:

1. extract questions from the image or text
2. read interview/questions.md
3. classify each question
4. edit interview/questions.md
5. report 新增 / 追加 / 跳过 / 未匹配