GrowReach v3 — Post Boost Comment Similarity Analysis
Date: July 14, 2026
Data source: post_boost_activities for boost_id=8 and boost_id=10, plus persona and comment_strategies data for all commenters
Author: Hermes Agent
Executive Summary
When multiple GrowReach users comment on the same LinkedIn post via Post Boost, their comments are visibly similar — same opener, same structure, same ending pattern, same length, same paragraph count. Anyone reading the post would immediately suspect AI automation. This report diagnoses the root causes using real data from 11 comments across 2 boosted posts, and outlines the fixes needed for v3.
1. The Data
Post Boost 8 — Post by Balraj Singh
Post topic: Engineering skills in the AI era (learn the work around the tools, not just the tools)
Comments: 5 (from 5 different users)
Post Boost 10 — Post by Puneet Jain
Post topic: Being a Product Manager in Tech (responsibility without authority, balancing constraints)
Comments: 6 (from 6 different users)
Commenters
| User | Persona Role | Company | Industry | Years Exp |
|---|---|---|---|---|
| Pranjal Aswani (48) | Engineer | Proximity Works | Tech | 7 |
| Johan Anthony (49) | Engineer | 10xers | Tech | 6 |
| Thiespo AKOMG (32) | Community manager | Beau-zen | Professional services | 5 |
| Ode Afdal (57) | Engineer | Growreach AI | Technologies | 10 |
| Gaurav Kushwaha (53) | Engineer | Thunes | FinTech | 13 |
| Rahul Aswani (36) | Founder | 10xers Labs | Software/AI Consulting | 12 |
| Pranjal Jain (61) | Engineer | CDAC | IT | 1 |
| Varun Sharma (60) | DevOps Engineer | Infosys | IT | 11 |
2. The Similarity Patterns
Aggregate Statistics (11 comments total)
| Pattern | Occurrence | Rate |
|---|---|---|
| Starts with “The point about…“ | 6 of 11 | 55% |
| Ends with a question mark | 10 of 11 | 91% |
| Ends with “How do/are/would you…“ | 9 of 11 | 82% |
| Exactly 3 paragraphs | 10 of 11 | 91% |
| Word count between 100-126 | 11 of 11 | 100% |
All gravitate to validate_extend structure | ~11 of 11 | ~100% |
Comment-by-Comment Breakdown
Post Boost 8 (Balraj’s post)
| # | Commenter | Opener (first 8 words) | Ends with Q? | ”How do you”? | Paras | Words |
|---|---|---|---|---|---|---|
| 1 | Pranjal Aswani | ”The point about protecting engineering standards…” | Yes | Yes | 3 | 112 |
| 2 | Johan Anthony | ”From an architecture perspective, the point about…” | Yes | Yes | 3 | 106 |
| 3 | Thiespo AKOMG | ”From a community management perspective, the point…” | No | No | 2 | 102 |
| 4 | Ode Afdal | ”The point about protecting engineering standards…” | Yes | Yes | 3 | 104 |
| 5 | Gaurav Kushwaha | ”The point about protecting engineering standards…” | Yes | Yes | 3 | 107 |
Post Boost 10 (Puneet’s post)
| # | Commenter | Opener (first 8 words) | Ends with Q? | ”How do you”? | Paras | Words |
|---|---|---|---|---|---|---|
| 1 | Pranjal Aswani | ”The point about arguing trade offs from…” | Yes | No | 3 | 104 |
| 2 | Rahul Aswani | ”The point about responsibility without authority…” | Yes | Yes | 3 | 107 |
| 3 | Johan Anthony | ”From an architecture perspective, the tension you…” | Yes | Yes | 3 | 126 |
| 4 | Pranjal Jain | ”In my work at CDAC, I’ve seen how…” | Yes | Yes | 3 | 123 |
| 5 | Varun Sharma | ”The point about responsibility without authority is where…” | Yes | Yes | 3 | 106 |
| 6 | Ode Afdal | ”In my work across Southeast Asia, I’ve noticed…” | Yes | Yes | 3 | 114 |
What a Reader Sees
If someone scrolls through comments on Balraj’s post, they see 5 comments where:
- 3 start with the exact same phrase “The point about protecting engineering standards is where…”
- 4 end with a “How do/are you…?” question
- 4 have exactly 3 paragraphs
- All are between 102-112 words
- All follow the validate_extend structure (agree + add angle + question)
This is a dead giveaway for AI-generated content.
3. Root Cause Analysis
Root Cause 1: “The point about” Opener Is Reinforced by the Prompt
The v2 system prompt explicitly bans “The point about…” as an opener:
BANNED PATTERNS: Starting with generic praise. Never open with “Great post” or “Love this” or “Spot on” followed by nothing.
But the structure definitions themselves use phrases that the AI copies:
STRUCTURE 1 opener guidance: “Reference something specific from the post directly. Do not start with ‘The point…’ or ‘Your point about…’. Start with the actual idea.”
The AI reads “The point about” in the structure description and uses it anyway. The ban is listed in a different section (ANTI-AI DETECTION RULES) and the AI does not connect the ban to the structure section. 55% of comments start with “The point about” despite the explicit ban.
Location: System prompt lines 25, 120-121
Root Cause 2: No Structure Coordination Across Commenters
When 5 different AI calls read the same post, they independently converge on validate_extend as the “best” structure. The system prompt says “Do not default to the same one every time” — but each AI call is independent. No AI knows what structures the other commenters already used.
User A → AI reads post → picks validate_extend → "The point about..."
User B → AI reads post → picks validate_extend → "The point about..."
User C → AI reads post → picks validate_extend → "The point about..."
User D → AI reads post → picks validate_extend → "The point about..."
User E → AI reads post → picks validate_extend → "The point about..."
This is a coordination problem, not a prompt quality problem. Even a perfect prompt cannot solve this if every AI independently picks the same structure for the same post.
Location: System prompt lines 17-27
Root Cause 3: “How do you…” Question Ending Is the Default
The system prompt provides 6 engagement invitation approaches:
- Bold statement that invites disagreement
- Share an experience that invites others to share theirs
- Genuine question
- Open reflection
- Humble admission
- Sometimes just end
But 82% of comments end with “How do/are/would you…” — the AI defaults to option 3 (genuine question) every time. The prompt says to vary, but provides no enforcement mechanism. The question format is the path of least resistance for the AI.
Location: System prompt lines 84-92
Root Cause 4: Paragraph Limit Ignored
The system prompt says “1 to 2 max” for most structures:
- validate_extend: “Paragraphs: 1 to 2 max”
- challenge_nuance: “Paragraphs: 1 to 2 max”
- share_experience: “Paragraphs: 1 to 2 max”
- ask_explore: “Paragraphs: 1 to 2 max”
But 91% of comments have exactly 3 paragraphs. The AI is consistently ignoring the paragraph limit. This makes every comment look the same shape.
Location: System prompt lines 22, 30, 40, 50
Root Cause 5: Persona Differences Change Content, Not Form
Looking at the persona data, the inputs ARE different across commenters:
| Variable | Variation across commenters |
|---|---|
| Role | Engineer, Community manager, Founder, DevOps Engineer |
| Company | Proximity Works, 10xers, Beau-zen, Growreach AI, Thunes, CDAC, Infosys |
| Industry | Tech, Professional services, Technologies, FinTech, IT |
| Years of experience | 1 to 13 |
| Voice brief | Different per user (detailed bios) |
But the AI uses these differences only to change what experience to cite in the comment body — not the shape of the comment. The form (opener, structure, paragraph count, ending pattern) converges because the system prompt’s structure examples dominate the output shape.
Key insight: 4 out of 5 commenters in Boost 8 are engineers in tech. The persona diversity is low. Even when personas differ (Thiespo is a community manager), the AI still forces the comment into the same validate_extend template.
Root Cause 6: Commenting Strategy Fields Are Mostly Empty
| User | content_tone | content_persona | content_priority | target_audience | content_goal | voice_brief |
|---|---|---|---|---|---|---|
| Pranjal Aswani | Professional | Industry Expert,Innovator | Authenticity first | Industry Peers… | Career Opps… | ✅ Detailed |
| Johan Anthony | (empty) | (empty) | (empty) | (empty) | (empty) | ✅ Detailed |
| Thiespo AKOMG | (empty) | (empty) | (empty) | (empty) | (empty) | ✅ Detailed |
| Ode Afdal | (empty) | (empty) | (empty) | (empty) | (empty) | ✅ Detailed |
| Gaurav Kushwaha | (empty) | (empty) | (empty) | (empty) | (empty) | ✅ Detailed |
| Rahul Aswani | Conversational | Data-Driven Analyst… | Metrics first | Executives… | Build Authority… | ✅ Detailed |
| Pranjal Jain | (empty) | (empty) | (empty) | (empty) | (empty) | ✅ Detailed |
| Varun Sharma | None | None | None | None | None | NULL |
Only 2 out of 8 users have commenting strategy fields filled. 6 users have empty/NULL content_tone, content_persona, content_priority, target_audience, and content_goal. One user (Varun) has a NULL voice_brief. When these fields are empty, the user prompt template renders them as empty, and the AI has less differentiation signal — increasing convergence.
4. The Input → Output Map
Here’s what actually happens when 5 users comment on the same post:
| Input Layer | Same or Different? | Impact on Output |
|---|---|---|
| System prompt (14K chars) | SAME for all | Dominates output shape |
| Post text | SAME for all | Same topic triggers same structure |
| Post author | SAME for all | No effect on variation |
| Model (Gemma 4) | SAME for all | Same biases, same defaults |
| Persona (name, role, company) | Different | Changes cited experience only |
| Voice brief | Different (but mostly detailed bios) | Changes content slightly |
| Commenting strategy fields | Mostly empty (6/8 users) | No differentiation signal |
| Structure selection | Independent per AI call | No coordination = convergence |
Result: The only things that vary are persona and voice brief. But the AI uses these to change what to say, not how to say it. The how (structure, opener, paragraphs, ending) is dictated by the system prompt and converges.
5. Proposed v3 Fixes
Fix A: Eliminate Copyable Example Phrases
Problem: The system prompt contains example sentences in structure definitions that the AI copies verbatim (“The point about…”, “Spot on about…”).
Fix: Remove all example sentences from structure definitions. Replace with abstract guidance only:
“Examples: ‘Spot on about the latency issue.’, ‘This maps to…’”- “Reference a specific idea from the post without using ‘The point about’ or ‘Your point about’ as an opener. Jump directly into the idea.”
Type: Prompt-only change
Fix B: Force Structure Diversity in Post Boost
Problem: Multiple AIs independently pick the same structure for the same post.
Fix: When generating comments for a post boost, pass the system a list of structures already used by other commenters on this post. The AI must pick a different structure. This requires a backend change.
{
"structures_already_used": ["validate_extend", "challenge_nuance"]
}The user prompt would include:
== STRUCTURES TO AVOID ==
The following comment structures have already been used by other commenters on this post. Do NOT use these. Pick a different structure.
- validate_extend
- challenge_nuance
Type: Backend change + prompt change
Fix C: Ban “How do you…” as Default Ending
Problem: 82% of comments end with “How do/are/would you…?”
Fix: Add to banned patterns:
- “Do not end with ‘How do you…’, ‘How are you…’, ‘How would you…’ as a question. These are the most overused AI comment endings.”
- Add a self-check instruction: “Before returning the JSON, verify your comment does NOT end with ‘How do/are/would you…’ If it does, rewrite the ending using a different engagement approach.”
Type: Prompt-only change
Fix D: Enforce Paragraph Limits
Problem: 91% of comments have 3 paragraphs despite the prompt saying “1 to 2 max.”
Fix: Add a post-generation self-check:
- “After writing your comment, count the paragraphs. If you have more than 2, combine paragraphs until you have at most 2. Do not exceed 2 paragraphs.”
Type: Prompt-only change
Fix E: Require Persona-Specific Anchoring
Problem: Persona differences change content but not form.
Fix: Add instruction:
- “Your comment must be anchored in a specific detail from the commenter’s voice brief that is unique to this person. Do not write a generic comment that could apply to any engineer. Reference a specific project, company, or experience from the voice brief.”
Type: Prompt-only change
Fix F: Handle Empty Commenting Strategy Fields
Problem: 6 of 8 users have empty commenting strategy fields. The user prompt template renders them as empty, reducing differentiation.
Fix: Backend should pass defaults when fields are empty, or the user prompt template should omit empty sections entirely rather than rendering “Tone: ” with nothing after it. This gives the AI less “empty space” signal.
Type: Backend change (minor)
6. Fix Priority Matrix
| Fix | Type | Effort | Impact | Priority |
|---|---|---|---|---|
| A. Remove copyable example phrases | Prompt-only | Low | High | P0 — Immediate |
| C. Ban “How do you…” ending | Prompt-only | Low | High | P0 — Immediate |
| D. Enforce paragraph limits | Prompt-only | Low | Medium | P0 — Immediate |
| E. Require persona-specific anchoring | Prompt-only | Low | Medium | P1 — Next |
| B. Force structure diversity | Backend + Prompt | Medium | Critical | P1 — Next |
| F. Handle empty strategy fields | Backend | Low | Medium | P2 — Later |
7. Conclusion
The v2 system prompt is well-designed for individual comment generation — when one user comments on one post, the output is natural and varied. But it was not designed for the Post Boost scenario where multiple users comment on the same post simultaneously.
The convergence is caused by:
- Copyable example phrases in the prompt (55% same opener)
- No coordination mechanism across commenters (100% same structure)
- Default to question endings (82% same ending)
- Ignored paragraph limits (91% same paragraph count)
- Empty commenting strategy fields reducing differentiation
v3 needs both prompt changes (Fixes A, C, D, E) and a backend change (Fix B) to pass structure diversity context. The prompt-only fixes can be deployed immediately. The backend fix for structure coordination is the highest-impact change for the post boost use case specifically.