interviewforge

name

interviewforge

description

Turn local interview videos, interview recordings, or interview audio files into local-first Chinese interview review PDFs focused on interviewer questions, the candidate's answers, concise sourced standard answers, compact coding/technical follow-up review, and short follow-up learning resources. Use when the user provides a local .mov/.mp4/.m4a/.wav recording and asks for 面试复盘, 面经整理, 面试官问题, 我的回答, 建议答案, 标准答法, 回答评价, 代码题复盘, 后续巩固资料, Q&A extraction, speaker/role inference, source-backed corrections, or a polished LaTeX/PDF report. The skill extracts audio locally, transcribes locally when possible, infers interviewer/candidate roles, registers local/public sources, renders a wdkns-style LaTeX report, and validates the final PDF with pdfinfo/pdftotext/pdffonts.

InterviewForge

Use this skill to convert any local interview video or audio recording into a Chinese interview review PDF. Optimize for: 面试官问题覆盖率 + 我的回答 > 简短建议答案 > 一句评价/关键扣分点 > 后续巩固资料.

The output is not a course note, not a raw transcript, and not a long training plan. Keep the report compact enough to review before the next interview.

Local-First Rule

Do not upload interview audio, video, transcripts, screenshots, resume details, or company-sensitive content unless the user explicitly asks for a cloud route.

Preferred ASR order:

1. whisperx with diarization when available.
2. mlx-whisper or faster-whisper for local ASR.
3. openai-whisper CLI fallback.
4. If diarization is unavailable, infer roles from dialogue structure and mark speaker confidence.

Privacy Defaults

Treat the skill package as reusable and non-personal:

• do not include real interview transcripts, company names, candidate names, usernames, home directories, or absolute local paths in bundled examples;
• do not show local absolute paths in the final PDF or references.md by default;
• record local evidence with event_id, artifact_id, basename-only paths, and time ranges;
• keep full paths only in transient internal commands or when the user explicitly asks for a debug/provenance bundle;
• remove extracted audio, LaTeX intermediates, temporary text extracts, and OS metadata files after validation.

Output Contract

Create one work directory per recording. Keep the final output clean and two-level:

interview_review.pdf
references.md
supporting_files/
  interview_review.tex
  review_plan.json
  source_registry.json
  question_candidates.json
  interview_events.json
  transcript_normalized.json
  transcript_raw.json
  transcript_raw.srt
  speaker_map.json
  media_probe.json
  run_manifest.json
  evidence_frames.json
  quality_report.json
  build_log.md
  <evidence-frame>.png

Keep Markdown output as a draft only. The final deliverable is <run-dir>/interview_review.pdf, with human-readable citations in <run-dir>/references.md. Put rebuild/provenance files in <run-dir>/supporting_files/. Do not leave user-facing output split across audio/, source/, logs/, analysis/, transcript/, report/, and frames/ directories.

Delete or avoid retaining bulky/transient files after the report is validated unless the user asks for them: extracted audio.wav, LaTeX .aux/.fls/.fdb_latexmk/.out/.toc/.xdv/.log, duplicate ASR .txt/.tsv/.vtt, .DS_Store, and temporary run directories.

Use natural report titles in metadata.title, such as <company/team/topic> 面试复盘. Keep metadata.subtitle empty by default unless the user asks for a specific subtitle. Do not expose internal style names, version labels, or pipeline labels in user-facing PDF titles.

Workflow

Use scripts/local_interview_pipeline.py for deterministic local steps:

SUPPORT=<run-dir>/supporting_files
python scripts/local_interview_pipeline.py init-run --workdir <run-dir> --input <video>
python scripts/local_interview_pipeline.py probe --input <video> --out-dir "$SUPPORT"
python scripts/local_interview_pipeline.py extract-audio --input <video> --audio "$SUPPORT/audio.wav"
python scripts/local_interview_pipeline.py transcribe-local --audio "$SUPPORT/audio.wav" --out-dir "$SUPPORT"
python scripts/local_interview_pipeline.py normalize-transcript --json "$SUPPORT/transcript_raw.json" --out-dir "$SUPPORT"
python scripts/local_interview_pipeline.py infer-speakers --transcript "$SUPPORT/transcript_normalized.json" --out-dir "$SUPPORT"
python scripts/local_interview_pipeline.py extract-question-candidates --transcript "$SUPPORT/transcript_normalized.json" --out-dir "$SUPPORT"
python scripts/local_interview_pipeline.py draft-review-plan --workdir <run-dir>
python scripts/local_interview_pipeline.py extract-frames --input <video> --events "$SUPPORT/interview_events.json" --out-dir "$SUPPORT"
python scripts/local_interview_pipeline.py enrich-real-answers --workdir <run-dir>
python scripts/local_interview_pipeline.py build-answer-polish-queue --workdir <run-dir>
# Agent/LLM reviews answer_polish_queue.json and writes cleaned questions/answers back to review_plan.json.
python scripts/render_interview_tex.py --review-plan "$SUPPORT/review_plan.json" --template assets/interview-review-template.tex --out "$SUPPORT/interview_review.tex"
python scripts/local_interview_pipeline.py compile-pdf --tex "$SUPPORT/interview_review.tex"
python scripts/local_interview_pipeline.py finalize-layout --workdir <run-dir>
python scripts/local_interview_pipeline.py validate-report --workdir <run-dir>

The agent owns the judgment-heavy parts:

• inspect transcript segments and evidence frames;
• inspect supporting_files/question_candidates.json;
• write supporting_files/interview_events.json;
• write supporting_files/source_registry.json;
• write or refine supporting_files/review_plan.json;
• rerun render/compile/validate until the report is useful and clean.

draft-review-plan is a high-recall scaffold step, not final judgment. It should create enough question cards from question_candidates.json and nearby transcript context so the agent can refine wording rather than starting from a 5-question summary. Review and merge noisy candidates, but keep substantive follow-ups as separate cards.

enrich-real-answers fills each question with a transcript-grounded answer window. build-answer-polish-queue creates the evidence package that an agent/LLM must review before writing question_cleaned and my_answer_cleaned. Do not treat raw ASR text as final prose or as a final interviewer question. question_cleaned should be the interviewer's concrete question in normal Chinese, close to the evidence window, and not a generic topic title. The cleaned answer must preserve the candidate's answer order and facts while fixing obvious ASR typos, terms, repetition, punctuation, and filler. Review low-confidence answers for ASR noise before final delivery.

Compact Sourced Report Standard

Set metadata.report_style to compact_sourced.

The PDF must contain these main sections:

1. 首页摘要
2. 面试官问题与我的回答
3. 重点追问复盘
4. 代码题复盘
5. 高风险技术点速记
6. 参考来源
7. 后续巩固资料

Do not add 7天训练计划, 证据与转写说明, or long raw transcript sections unless the user explicitly asks.

Each question card should include:

• 时间
• 面试官问题
• 我的回答
• 建议答案
• 来源
• optional 一句评价

面试官问题 must use question_cleaned, not the raw ASR fragment. Fix ASR spelling and missing words only when supported by the nearby transcript and answer window; if the exact question is unrecoverable, write a bounded form such as 面试官追问：这里的 X 具体指什么？ instead of rendering nonsense.

我的回答 must be a cleaned version of the candidate's real answer from the transcript window, not a one-sentence evaluation. Keep the original order and concrete details; remove only obvious ASR hallucinations, long repeated phrases, filler, and screen-share noise. Use answer_summary only as an internal short summary or overview signal.

Each question card should show a compact top-left quality tag like passable 3/5, risky 2/5, or strong 5/5, plus an importance tag like key or covered. The score is a review aid based on answer completeness, clarity, and ASR confidence; it is not an absolute interview outcome.

Keep 建议答案 short. Use 1-3 sentences unless the user asks for deeper teaching notes.

Coverage rules:

• Preserve most substantive interviewer questions and follow-ups. Do not collapse a long technical discussion into one broad card.
• Only filter greetings, logistics, ASR hallucinations, and repeated acknowledgements.
• For interviews longer than 30 minutes, default targets are: 10-16 question cards for <=35 minute interviews, 14-20 for 35-65 minute interviews, and 18-28 for longer interviews. Actual count should follow transcript evidence.
• Mark 5-8 high-risk/high-value questions with is_key: true; those receive fuller review in 重点追问复盘.
• Non-key questions can use 1-2 sentence 建议答案; key questions should get 2-5 sentences.

Add learning_resources for 后续巩固资料 at the end of the report:

• include 5-8 items, grouped by topics revealed in this interview;
• each item needs topic, title, type, url or citation, why, and priority;
• optional fields: language, duration;
• use traceable sources: papers, official docs, official repositories/docs, university/course notes, original blogs, Zhihu articles, YouTube videos, Bilibili videos, or comparable knowledge-sharing pages;
• keep each why to one concise sentence explaining the exact gap it helps fix;
• do not invent links or source titles. If reliable material is unavailable, write 来源不足，建议补材料 in the report rather than fabricating a resource.

Source Rules

Use supporting_files/source_registry.json and mirror it into review_plan.json as source_registry.

• Interview facts must cite local sources: transcript_normalized.json, interview_events.json, or evidence frames.
• Technical corrections must cite public traceable sources: papers, official docs, official repositories/docs, course notes, or original blog posts.
• Project ownership, exact work done, experimental numbers, and company-specific claims must cite local video/transcript or user-provided materials. If not sourced, phrase as 建议表达方式, not as a fact.
• If no reliable source exists, the answer must say 来源不足，建议补材料. Do not invent citations.
• Use short refs in body, e.g. [L-q03], [PPO], [DAPO].
• Follow-up learning resources are not evidence for what happened in the interview. They should be selected after reading the current transcript/events and should not be hard-coded for a company, team, or candidate.

Writing Rules

• Write in Chinese unless the user requests another language.
• Keep interviewer questions concrete and close to the original wording.
• Filter greetings, repeated acknowledgements, screen-share logistics, silence hallucinations, and ordinary small talk unless they affect interview signal.
• Evaluate as an interviewer would, but keep the critique short.
• For high-risk technical points, include only the formula/rule needed to fix the answer.
• For code, prefer a compact interview-writeable version, invariant, complexity, and next-time oral script.
• Mark uncertainty rather than pretending diarization or ASR is perfect.

Visual Evidence Rules

Use screenshots only when they carry evidence value:

• code editor or judge result;
• whiteboard or shared-screen formula;
• architecture diagram;
• table/plot/benchmark;
• system design sketch.

Do not insert random face frames or decorative images. Every included image must have a time provenance footnote on the same page. Use frames/evidence_frames.json to record path, timestamp, caption, and reason.

Bundled Resources

• assets/interview-review-template.tex: wdkns-style LaTeX template with verdictbox, riskbox, betterbox, evidencebox, drillbox, and codebox.
• scripts/render_interview_tex.py: render review_plan.json into complete LaTeX. It supports both legacy V2 and compact_sourced.
• scripts/validate_report.py: validate PDF existence, extractable text, compact headings, source backing, timestamps, placeholders, image paths, fonts, and page metadata.
• scripts/local_interview_pipeline.py: local probing, audio extraction, transcript normalization, role inference, frame extraction, LaTeX compilation, validation orchestration.
• references/data-contracts.md: JSON contracts for transcript, events, source registry, and review plan.
• references/report-rubric.md: scoring dimensions and interviewer-signal interpretation.
• references/formulas/ppo_grpo_clipping.md: PPO/GRPO/DAPO clipping formula card.
• references/formulas/kl_direction_cheatsheet.md: KL direction and mode/mean-seeking cheatsheet.
• references/answer_rewrites/project_claim_defense.md: project claim defense templates.
• examples/minimal/: non-sensitive transcript -> review plan -> PDF smoke-test sample.

Final Checklist

Before delivery:

• compile with latexmk -xelatex -interaction=nonstopmode -halt-on-error;
• run pdfinfo, pdffonts, and pdftotext;
• run scripts/validate_report.py --workdir <run-dir>;
• ensure extracted text contains 首页摘要, 面试官问题与我的回答, 重点追问复盘, 代码题复盘, 高风险技术点速记, 参考来源, 后续巩固资料;
• ensure extracted text does not contain 7天训练计划 or 证据与转写说明;
• fail on placeholders such as TODO, TBD, [在此填写], <<...>>;
• check every question has 建议答案 and at least one source ref;
• check question coverage against question_candidates.json / interview_events.json;
• check public sources have URL/citation and local sources have path/event id;
• check learning_resources contains 5-8 useful traceable items;
• check all \includegraphics{...} paths exist;
• record commands, fallbacks, and limitations in supporting_files/build_log.md.