Turn the last few weeks of interview write-ups on LeetCode Discuss into one doc in Traditional Chinese: the posts worth reading, the questions that came up more than once, and which of those problems this repo has never practised.
Invocation: /lc-discussion <keywords> [--days N] — e.g. /lc-discussion google,
/lc-discussion google l3 --days 30, /lc-discussion amazon oa.
Why this exists
leetcode.com/discuss/ renders in the browser, so fetching the page returns an
empty shell, and most of what the listing shows is noise: team-matching threads,
pay questions, roadmap posts that mention the company name once. In the Oct 2026
Google scan, about 30 of 200 posts described an actual coding question. The
useful part is reading those 30 and lining them up against the log. That turned
up five problems that had come straight from L3/L4 loops and had never been
practised (1235, 3026, 1218, 2812, 1101).
script/lc_discussion.py
does the mechanical part. It calls the GraphQL endpoint the page itself uses
(no login), keeps only posts that carry every keyword term, fetches the full
markdown of the ones you pick, and cross-references LC numbers against README and
data/progress.txt offline, using the parsers script/l3_core.py uses.
Prime directives
- First-hand beats aggregated. A post saying "I was asked X" counts. A post saying "the 8 problems Google asks in 2026" is pattern-level, often built from Glassdoor or Blind, and is reported with that caveat. Never treat it as a sighting.
- Name or match, never guess silently. An LC number is named when the post
links it or writes it. Anything else is a match you inferred from the story
(the mouse-and-cat grid → LC 2812), and the doc labels it
相近. A question with no good match is listed as無對應 LC. Do not force a number onto it. - The log is the record. Every LC number in the doc gets the
xrefresult: the log's latest verdict first, README's status cell beside it. A problem is never described as solid because README saysOK. - Summarise, don't republish. Each post gets a link and a one-line digest of the question. Quote a constraint or an example when the question depends on it, never a whole post.
- Read-only on the record. This skill writes one file under
doc/ws/. It never writes README ordata/progress.txt. Drills it suggests are logged with/lc-log.
The steps
1. Search
python3 script/lc_discussion.py search google --days 90 --summary
python3 script/lc_discussion.py search google l3 --days 30
Each line gives the date, upvotes, topic id, title and URL. Every keyword term
must appear in the title or summary; --loose drops that filter, and
--order MOST_RELEVANT ranks by match instead of date. --pages (default 4
× 50) widens the sweep. --json FILE keeps the list for later.
2. Shortlist
Keep the posts that describe a question: interview experiences, phone screens, onsites, OAs, compiled question lists. Drop team matching, pay, timelines, "should I accept", and anything about another role family (SRE, FDE, data centre) unless the keywords asked for it. Put the target level first when there is one, and say how many posts were scanned and how many were kept.
3. Read
python3 script/lc_discussion.py fetch 8543506 8527382 8524882
The full markdown of each post, with a named: line listing any LC links or
numbers it contains. For more than a handful, add --dir <scratch dir> and read
the files from there; --dir never points inside the repo. Note each question,
its follow-ups and, if the post gives one, the verdict the round got.
4. Map to LC numbers
For each question, write down the LC number and whether it is named or 相近
(directive 2). Count repeats across different posts: a compiled list that
copies another post is not a second sighting. Patterns that repeat (the same tree
DP from two candidates, three variants of a rate limiter) are the main finding
of the doc.
5. Cross-reference
python3 script/lc_discussion.py xref 1235 3026 963 359 1101 2812
Offline. For each number it prints README's title and status, the log's latest
verdict (ok, again, none, or never), when it was last attempted, how many
attempts there were, and whether it is in the L3 core set. Sort the doc's
problems into three groups: never logged, again or no verdict, and
ok.
6. Write the doc
Write doc/ws/lc_discussion_<keywords>_<YYYYMMDD>.md (keywords joined with _,
the date the scan ran) in Traditional Chinese with this shape:
# LeetCode Discuss 掃描:<keywords>(<YYYY-MM-DD>)
> 掃描範圍:<N> 篇貼文(<from> – <to>),保留 <M> 篇描述實際題目的貼文。
## 值得讀的貼文
| 日期 | 貼文 | 內容 |
## 重複出現的題型
- ...
## 對照你的練習紀錄
### 從沒記錄過
| LC | 題目 | 出處 |
### 有紀錄但最新結果是 again 或沒有結果
### 已經 ok
## 建議的下一輪練習
House rules: LC titles, API names and code stay in English; every post is a
link; every LC number has its xref status; mark 相近 matches; tag every code
fence. doc/ws/ is not built into the site, so nothing else needs to change.
7. Report
Tell the user the scan size and how many posts were kept, the two or three
patterns that repeated, and the five never-logged problems you would drill first,
each with the post it came from. Give the doc's path. If they drill them, the
next step is /lc-log.
Do not
- ❌ count a compiled or "confirmed sightings" list as a first-hand report (directive 1)
- ❌ give a question an LC number without saying whether it was named or matched (directive 2)
- ❌ report a problem's state from README alone;
xrefthe log (directive 3) - ❌ paste whole posts into the doc; link them and summarise (directive 4)
- ❌ write
data/progress.txtor README, or save fetched posts inside the repo (directive 5) - ❌ scrape the HTML page or send a login cookie; the public GraphQL endpoint is enough
- ❌ commit or push unless asked
Worked example
/lc-discussion google on 2026-10-02, which produced
doc/ws/lc_discussion_google_20261002.md:
| Step | What it produced |
|---|---|
| 1 | search google --days 90 → 200 posts fetched, about 125 about Google interviews |
| 2 | about 30 describe a coding question; 8 kept as worth reading, with L3 first |
| 3 | fetch on those 30; the 75-question compilation (8525983) read as a list, not as sightings |
| 4 | named: 1235, 3026, 963, 359, 2402, 68, 60, 351, 1970; matched (相近): 1101, 2812, 1218, 249, 1254, 3453; repeats across posts: 1235 (2 onsites), the leaf-cutting tree DP (2), the right-only grid path count (2), stream variants of the rate limiter (3) |
| 5 | xref → never logged: 1235, 3026, 1101, 2812, 1102, 1218, 1970, 1293, 381, 68; again or no verdict: 963, 939, 1254, 2402, 1631, 621, 329, 410, 128, 91; ok: 200, 207, 994, 127 |
| 6 | the doc, in Traditional Chinese |
| 7 | suggested drill: 1235, 3026, 1218, 2812, 1101 |