GrowReach v3 — Activities vs Post Boost Comparison
Date: July 14, 2026
Data: 60 rows from activities table (auto-commenting flow) vs 11 comments from post_boost_activities (post boost flow)
Executive Summary
The activities table (auto-commenting) and post_boost_activities table tell two very different stories. The auto-commenting engine produces relatively varied comments — different strategies, varied openers, reasonable word counts, mostly 1-2 paragraphs. The post boost flow produces comments that are uniformly identical — same structure, same opener, same ending, same paragraph count, same length.
This suggests the problem may be specific to the post boost flow, not the prompt itself.
Side-by-Side Comparison
| Metric | Activities (Auto-Comment) | Post Boost | Verdict |
|---|---|---|---|
| Starts with “The point about” | 0% (0/60) | 55% (6/11) | Post boost has a unique opener bug |
| Starts with “The” | 72% (43/60) | 55% (6/11) | Both overuse “The” as opener, but activities varies the rest |
| Ends with question mark | 47% (28/60) | 91% (10/11) | Post boost almost always ends with question |
| Ends with “How do you…“ | 0% (0/60) | 82% (9/11) | Post boost has a severe “How do you” pattern |
| 3 paragraphs | 10% (6/60) | 91% (10/11) | Post boost ignores paragraph limit; activities mostly respects it |
| Avg word count | 84 | 110 | Post boost comments are significantly longer |
| Strategy variety | 5 of 6 structures used | ~1 (validate_extend only) | Post boost has zero strategy diversity |
| Opener variety (first 4 words) | 58 of 60 unique | 3 of 11 unique | Post boost openers are near-identical |
| voice_profile logged | 60/60 | 0/11 | Post boost doesn’t log metadata |
| strategy_used logged | 60/60 | 0/11 | Post boost doesn’t log metadata |
| ai_analysis populated | 0/60 | N/A | Not populated in either flow |
Strategy Distribution Comparison
| Strategy | Activities (60 comments) | Post Boost (11 comments) |
|---|---|---|
| validate_extend | 25 (42%) | 11 (~100%) |
| ask_explore | 21 (35%) | 0 (0%) |
| challenge_nuance | 7 (12%) | 0 (0%) |
| direct_observation | 5 (8%) | 0 (0%) |
| casual_react | 2 (3%) | 0 (0%) |
| share_experience | 0 (0%) | 0 (0%) |
The auto-commenting flow uses 5 of 6 structures with reasonable distribution. The post boost flow uses only validate_extend for every single comment.
Key Findings
Finding 1: The Prompt Works for Auto-Commenting
The v2 system prompt is performing reasonably well in the auto-commenting flow:
- 5 of 6 structures are used (only
share_experienceis unused) - Openers vary across 58 of 60 comments (first 4 words are unique)
- Paragraph count is mostly 1-2 (50% have 2, 40% have 1, only 10% have 3)
- Question endings are 47% (not 91%)
- “How do you…” endings: 0%
- Average word count is 84 (within the system prompt’s target ranges)
The prompt is NOT broken for the primary use case.
Finding 2: Post Boost Has a Different — and Worse — Problem
The same prompt produces dramatically different output patterns in the post boost flow:
| Symptom | Activities | Post Boost | Root Cause |
|---|---|---|---|
| All comments use validate_extend | No (5 structures used) | Yes (~100%) | No coordination — same post triggers same structure decision |
| ”The point about” opener | 0% | 55% | Multiple AIs reading the same post converge on the same opener |
| ”How do you…” ending | 0% | 82% | The same post triggers the same engagement approach |
| 3 paragraphs | 10% | 91% | The same post leads to same paragraph structure |
| Longer comments (avg 110 vs 84) | No | Yes | validate_extend’s 40-80 word range is being exceeded |
Finding 3: “The” Opener Is a Persistent Issue in Both Flows
Even in the activities table, 72% of comments start with “The”. The system prompt’s structure examples all start with “The” or similar determiners:
- “The real problem here isn’t…”
- “The focus on building…”
- “The computing power excuse…”
- “The shift toward…”
- “The gap between…”
This is less of a problem in auto-commenting because each post is different, so “The [different topic]” creates variety. But in post boost, all commenters read the same post, so they all extract the same key phrase and produce “The point about [same phrase]…”
Finding 4: Post Boost Missing Metadata
The post_boost_activities table doesn’t log strategy_used, voice_profile, word_count, or ai_analysis. This means:
- We can’t verify which structure the AI actually chose
- We can’t track voice consistency
- We can’t measure word count compliance
- We can’t do post-hoc analysis without manually reading comments
Rahul has already flagged this to the dev team to fix.
Finding 5: voice_profile Shows One Dominant User
In the activities data, 38 of 60 comments are from user 36 (Rahul), all with voice_profile “senior AI founder, direct and analytical.” This user has fully configured commenting strategies. The other 3 users (47, 48, 51) have fewer comments but also show variety in strategies and openers.
The auto-commenting data is skewed toward one user, but the pattern analysis still holds — even this single user’s 38 comments show opener variety and strategy diversity.
Root Cause: The Convergence Problem
The fundamental issue is not the prompt — it’s the lack of coordination when multiple AIs comment on the same post independently.
AUTO-COMMENTING FLOW (works well):
User A → AI reads Post 1 → picks ask_explore → "The real problem here..."
User A → AI reads Post 2 → picks validate_extend → "The focus on building..."
User A → AI reads Post 3 → picks challenge_nuance → "The computing power excuse..."
→ Different posts = different structures = different openers = natural variety
POST BOOST FLOW (broken):
User A → AI reads Post X → picks validate_extend → "The point about..."
User B → AI reads Post X → picks validate_extend → "The point about..."
User C → AI reads Post X → picks validate_extend → "The point about..."
→ Same post = same structure = same opener = obvious AI pattern
The prompt says “vary your choice” but each AI call is independent and doesn’t know what others chose. When reading the same post, the AI converges on the same “best” interpretation.
Implications for v3
| Fix | Needed for Auto-Comment? | Needed for Post Boost? | Approach |
|---|---|---|---|
| A. Remove copyable example phrases | Yes (reduce “The” openers) | Critical | Prompt-only |
| B. Force structure diversity | No | Critical | Backend + Prompt |
| C. Ban “How do you…” ending | No (already 0%) | Critical | Prompt-only |
| D. Enforce paragraph limits | Minor (already 90% compliant) | Critical | Prompt-only |
| E. Require persona-specific anchoring | Nice to have | Helpful | Prompt-only |
| F. Handle empty strategy fields | Nice to have | Helpful | Backend |
| G. Log metadata in post_boost_activities | No | Critical | Backend |
The Big Insight
The v2 prompt is 80% good. The auto-commenting flow proves it. The post boost problem is primarily a coordination problem that needs a backend fix (passing “structures already used” to each subsequent AI call), not just a prompt rewrite.
The prompt-only fixes (A, C, D) will help reduce the “The” opener pattern and improve the post boost output, but Fix B (structure coordination) is the only way to fully solve the post boost convergence problem.
Next Steps
- Immediate (prompt-only): Write v3 system prompt with fixes A, C, D, E — removes copyable phrases, bans “How do you” endings, enforces paragraph limits, requires persona anchoring
- Short-term (backend): Add
structures_already_usedparameter to the post boost comment generation flow so each subsequent commenter is forced to use a different structure - Short-term (backend): Add
strategy_used,voice_profile,word_count,ai_analysiscolumns topost_boost_activitiestable for observability - Test: Generate comments for the same post with v3 prompt + structure coordination and measure the improvement against the current 55% same-opener / 91% same-ending / 100% same-structure baseline