Turn an interview that "scattered as soon as it ended" into a reusable asset. Core deliverables:A Feishu retrospective document, including question-by-question diagnosis, 5-dimension scoring, better answer templates, and an action checklist for the next round.
Applicable
- You have already finished an interview and have the Q&Atext record(pasted / Feishu document link)
- You want to know which questions fell apart, why, and what a better solution looks like
- You want to prepare in a targeted way for the next round (next round at the same company / similar role)
Not applicable
- Haven't interviewed yet, want to optimize your resume → use
resume-optimizer/cv-expert - Only audio/screen recording and no text → first use
lark-minutes(Feishu Minutes) to transcribe, then return to this skill - Want to practice mock interviews (real-time Q&A) → this skill is post-event retrospective, not a practice partner
Input materials
| Material | Required | Description |
|---|---|---|
| Interview Q&A text record | ✅ | Paste directly; or provide a Feishu document link → use lark-doc to read |
| Target JD | Strongly recommended | To judge "what should be tested" + "whether your selling points came through" |
| Your own resume | Strongly recommended | To see whether answers and resume are self-consistent and whether stronger stories can be mined |
| Basic information | Guide completion | Company / role / round (HR/technical/director/final HR interview) / overall feeling / questions where you felt stuck |
The rougher the record, the better (fragmented recollections are fine), butdo not fabricate Q&A that did not happen—mark questions you cannot remember as "cannot remember"; retrospect why you cannot remember it, as that itself is also information.
Process
Phase 0 · Collect and confirm
- Get the text record. If it is a Feishu link → call
lark-docto read the body. - Ask for all basic information (company / role / round / overall feeling); when JD and resume are missing, proactively ask "Do you have the JD and resume? They can make the retrospective more accurate",use them if available; you can continue without them。
- When unsure about priority / scope, ask first; don't bury your head and analyze a huge amount.
Phase 1 · Reconstruct timeline
Split the record into an ordered "question list". Extract each question into a four-tuple:
- Question — what the interviewer asked
- Intent — what this question is really testing (knowledge point / ability to verify / whether it is a stress question or a hidden trap)
- My answer — how the user answered at the time (truthfully, without beautifying)
- State at the time — fluent / stuck / off-topic / collapsed under follow-up / cannot remember
Output a timeline table for the user to confirm, and only after confirmation move into question-by-question review—avoid doing a ton of analysis on a wrong reconstruction。
Phase 2 · Question-by-question review (core)
For each question, provide:
- Score — use the
references/rubric.mdsingle-question four-tier scale (Excellent / Good / Fair / Poor), with a one-sentence reason for the rating - Diagnosis — where the gap is: not understanding it, understanding but not explaining it clearly, scattered structure (no STAR), or not hitting the assessment points / not echoing the JD
- Better answer exemplar — give a version of the answer that can be practiced by following it. Dig for stronger stories from the resume, and align with the JD's assessment points. The exemplar mustsound like something a real person could say, not a list of knowledge points.
For questions where you fell apart under follow-up questions, mark the follow-up chain separately, and review "at which follow-up question you started being unable to handle it."
Phase 3 · Horizontal profile
Based on the whole interview, give 5-dimension scores (see the scale in references/rubric.md), with a one-sentence comment for each dimension:
- Professional/technical depth — whether you can withstand hardcore questions and whether there is depth
- Expression structure (STAR) — whether there is structure, or whether you just say whatever comes to mind
- Role fit — whether the answers align with the JD and bring out your selling points
- On-the-spot reaction — how you handle follow-ups, pressure questions, and questions you haven't prepared for
- Quality of questions you ask — whether you asked anything valuable in the final questions-to-interviewer segment / or actually lost points instead
Phase 4 · Consolidation
- Top 3 improvement actions — the three things most in need of change, specific enough to answer "what to do next time," not correct but empty platitudes
- Next-round preparation checklist — for the next round at the same company / similar roles, list the questions and areas to focus on preparing
- Question backlog list — the questions exposed in this interview that were answered poorly or could not be answered, listed as questions you can take away to practice
Output → Feishu review document
Use the lark-doc skill to create it (docs +create --api-version v2, with the body in Markdown).
- Title format：
面试复盘 · {公司}-{岗位}-{轮次} · {YYYY-MM-DD}(use today's date) - Body structurestrictly follow
references/report-template.md - After creating it, send the document link back to the user
Before writing the document, first walk the user through the complete content in the conversation,and only create the Feishu document after the user confirms(consistent with the habit of "discuss first, then act").
Style constraints
- Honest, no chicken soup. If the answer is bad, say it's bad, but give "how to make it better."
- The better answer exemplar is the core value of this skill—it must be specific, something you can say by following it, fit the user's real background, and not be a generic template.
- Everything is based on the real records provided by the user; do not fabricate for the user questions the interviewer did not ask or answers that did not happen.