| name | chat-export-report |
| description | Use when user provides exported chat history (WeChat / Telegram / iMessage / QQ exports as .md / .txt / .json) and asks "我和 X 聊了啥"、"看完整个 md"、"细说 XX"、"是不是有 Y"、"what did I and X talk about", wants topical breakdown, requests detail on specific themes (relationships / work / health / events), or seeks honest interpretation of conversation dynamics. Triggers on large chat dumps (>1000 lines) where direct full read is impractical.
|
| metadata | {"author":"innei","version":"0.1.0"} |
chat-export-report
Analyze massive exported chat logs (thousands to hundreds of thousands of lines) and produce a layered, drill-downable report grounded in original quotes.
When to invoke
- User provides path to exported chat (WeChat / Telegram / iMessage / QQ; .md / .txt / .json)
- File far exceeds a single
Read window (>2000 lines)
- User asks "what did we talk about", "read the whole file", "tell me more about X", "was there Y between us"
- User wants an honest read of relationship dynamics, emotional tone, or missed opportunities
Do not use for:
- Single messages or short conversations
- Looking up one specific fact (just
grep)
- Full line-by-line translation or re-export
Reading strategy
Step 1 — Measure size
wc -l <file>
Step 2 — Three-zone sampling
Never read the whole file at once. Establish baseline tone first:
| Zone | offset | limit |
|---|
| Opening | 1 | 200 |
| Middle | total/3 and total*2/3 | 120–150 each |
| Tail | total-200 | 200 |
Four to five samples fix the time span, message density, opening / closing state, and any pause points.
Step 3 — Topical grep
Sweep for line numbers by topic family, then Read surrounding context as needed:
| Topic | Keyword family |
|---|
| Work | 实习|入职|面试|mentor|论文|三方|加班|KPI|大厂|interview|onboard |
| Food | 外卖|做饭|盒饭|吃饭|饿|食堂|takeout|cook |
| Housing | 租|合租|中介|宿舍|房租|老家|rent|roommate |
| Romance | 对象|男朋友|女朋友|喜欢|分手|相亲|介绍|追|date|crush |
| Health | 抑郁|emo|失眠|噩梦|结节|医院|体检|depress|insomnia |
| Current events | 封|核酸|疫情|户口|历史|lockdown|covid |
| Hobbies | 游戏|番|二次元|新海诚|动漫|追剧|game|anime |
| Sexuality | 性取向|gay|les|喜欢男|喜欢女 |
Per family, take 30–50 line numbers; then Read 60–150 lines of surrounding context to verify.
Output layering
Reply in four layers, deepest only on follow-up — never dump all four at once, leave room for the user to drill down:
L1 — Overview (500–800 words)
Time span, total message count, theme list (4–7 buckets), overall tone.
L2 — Theme breakdown (one section per bucket, three to five bullets, sparse direct quotes)
Use the topic families from Step 3 as section headings; 100–200 words each.
L3 — Single-theme deep dive (800–1500 words)
Triggered when user says "tell me more about X" / "细说 XX". Include sub-sections, original quotes, dates.
L4 — Subjective judgment (relationship / opportunity / missed chance)
Triggered by "was there a chance", "did I miss it", "我是不是错过". Three-part structure is mandatory:
- Evidence for "yes" (positive signals, usually fewer)
- Evidence for "no" (negative signals, usually more)
- Lean + reasoning (no emotion, just the objective read)
Quoting discipline
- Every quoted line carries a date
YYYY-MM-DD
- Strictly identify the speaker: lines prefixed with
**Innei**: are the user; unprefixed lines are the other party
- Never fabricate quotes. Unsupported claims may be flagged as "impression" or "lean" — never disguised as evidence
- Keep quotes short (1–2 lines); paraphrase longer passages with line-range citation
Honest interpretation
When the user asks a subjective question, do not pander to their hoped-for answer. Read by objective signals:
| Signal | Meaning |
|---|
| Other party repeatedly offers to "matchmake" / "introduce someone" | They've placed themselves outside the candidate set |
| Explicit "I don't need a partner" | Single-life declaration |
| Invitation (meet / dinner) declined without leaving an opening | Hard boundary |
| Cold-shouldering emotion ("emo will make you ugly", "stop calling yourself depressed") | Refusal to be drawn in deeply |
| Geographic / career divergence ("I'll head home in two years") | Long-term incompatibility |
| Sexuality teasing from a "elder-sibling" register ("don't lock yourself into one gender") | Joke, not test |
| User self-talks themselves out ("I never make the first move", "I'm doomed to be alone") | Self-foreclosure |
Reading by these signals yields a relatively objective lean, free of the user's memory-tinted lens.
Common mistakes
| Wrong | Right |
|---|
| Read the entire file in one go | Sample + grep |
| Judge from opening alone | Tail matters too — it has the latest state |
| Claims without quotes | Always cite date + original line |
| Mirror the user's hopes | Two-sided reading, objective lean |
| Confuse speakers | Check prefix every time before quoting |
| Ignore time | Note start–end dates per section |
| Trust user-supplied path blindly | find / ls to verify first |
cat a giant file | Use Read or head/sed slicing |
| Dump L1+L2+L3+L4 in one reply | Stop at L1/L2; wait for follow-up |
| Misattribute or invent quotes | Strict honesty — omit rather than fake |
Red flags
- Writing "they probably talked about X" without a quote → STOP, go grep
- Quote without a date → STOP, add it
- Subjective judgment with only one side → STOP, add the counter-side
- Answering "yes" or "no" to a "was there Y" question without listed evidence → STOP, list signals
