Comment Sentiment Analysis
Cut the launch-day comment stream into four tiers: "Must reply immediately" / "Must reply carefully" / "Record" / "No need to reply." P0 risks must be escalated to the developer within 30 minutes. On ProductHunt, first-day comment response speed directly affects ranking – slow recognition = ranking loss.
What this skill does not do: write reply content (that's for reply-drafting), execute auto-replies, or delete any comments.
Meta Information
| Field | Value |
|---|---|
| ID | de_dev_02_sk07 |
| Type | atomic |
| Employee | Launch Commander |
| Capability Family | Detection & Judgment · Sentiment Classification & Risk Alert |
| Complexity | M |
| Stage | P1 |
| Reusability | Medium (T+1 postmortem reuses comment data) |
| Asset Form | Deep prompt + Python script (no model calls, no API key) |
Capability Description
- Five Sentiment Labels: 🟢 Positive / ⚠️ Negative / ⚪ Neutral / 🔵 Question / 🔴 Risk, with negation flip and "feature request ≠ positive review" correction
- Six Risk Subtypes: Vote manipulation accusations, security/privacy concerns, refund requests, plagiarism/competitor claims, product viability doubts, paid review allegations – each with standard handling actions and SLA
- Priority Queue: P0 Risk (30 minutes) → P1 Negative (PH first-day 30 minutes) → P2 Question (2 hours) → P3 Positive (24 hours) → P4 Neutral (48 hours)
- Health Baseline: Positive ratio ≥ 60% and risk = 0 is healthy; negative > 20% triggers pinned clarification and coordinated response; questions > 30% indicates gaps in launch materials
Input Specification
| Field | Type | Required | Description |
|---|---|---|---|
comments | array | ✅ | Comment stream, each with text (try to include full text) + user + platform |
platform | string | ⬜ | Main platform: ProductHunt / Reddit / V2EX / X / Jike / Weibo / WeChat Official / Generic |
product | string | ⬜ | Product name, used for context |
Output Specification
summary: Five-class count, positive ratio, P0/P1 count, health line determinationqueue: Priority queue (P0→P4), each item with risk subtype / SLA / suggested action / pending manual confirmation status
Artifacts:
| File | Content |
|---|---|
out/comment_classification_list.xlsx | Priority queue (risk rows highlighted red) / classification details / summary |
out/sentiment_distribution.png | Pie chart of five sentiment categories |
out/emotion_scan.json | Machine-readable results, for workflow (daily-comment-duty-flow) consumption |
Usage Steps
Method 1: Pure Prompt (Fastest, Includes Context Judgment)
- Paste the entire contents of
prompt.txtas the system prompt - Provide the comment stream (full text, with platform labels)
- Receive sentiment overview + priority queue + items requiring model review (sarcasm, fake accounts)
Method 2: With Script (Precise Word Scoring + Excel/PNG Output)
python3 <SKILL_DIR>/scripts/emotion_scan.py --input input.json --outdir out
python3 <SKILL_DIR>/scripts/emotion_scan.py --demo # Works even without input
Division of Labor: The script handles wordlist scoring, negation flips, risk rule scanning, and P0-P4 queuing (complete, no guessing); the model handles sarcasm detection, fake account review, and veteran user negative comment attribution (things the script cannot do). The two results are cross-checked.
Boundaries (What This Skill Does Not Do)
- ❌ Does not delete comments or recommend deletion – "Acknowledge + offer solution" is the only correct posture for negative comments
- ❌ Does not write reply content (that's reply-drafting's job), only provides handling direction
- ❌ Does not count feature requests as positive reviews (would inflate sentiment and misjudge PMF)
- ❌ No automated context judgment – sarcasm and fake accounts must be reviewed by the model
- ❌ If the comment stream is empty, outputs "Launch day 0 comments" and prompts to check the launch link – does not fabricate data
Example Call
Input (examples/input.json, excerpt):
{
"product": "ClipMate (Clipboard History Tool)",
"comments": [
{"user": "maker_fan99", "platform": "ProductHunt", "text": "Congrats! Finally a clipboard manager that doesn't feel bloated. Love the search speed."},
{"user": "angry_bob", "platform": "ProductHunt", "text": "Crashed twice within 10 minutes. How do I get a refund?"},
{"user": "privacy_nut", "platform": "V2EX", "text": "Will clipboard data be uploaded to the server?"}
]
}
Output (script run, exit code 0): Positive 1 / Negative 0 / Neutral 2 / Question 1 / Risk 3, P0+P1 total 3 items – angry_bob "refund" hits refund request P0; "doesn't feel bloated" is correctly identified as a positive review after negation flip. See examples/output.md for details.
Workflows
daily-comment-duty-flow(daily comment duty, every 15 minutes)launch-battle-review-flow(launch postmortem review, T+1)
Compliance Statement
- Outputs are labeled "AI-generated content"; all actions require "pending manual confirmation" before execution
- Comment content is user data: the classification list is limited to launch duty use only; not shared externally, not publicly screenshot (positive quote citations require permission)
- No auto-replies; all reply drafts must go through reply-drafting and receive manual confirmation
This skill follows the bangwozuo Digital Employee Asset Specification v3.0