Building a Reddit account network with AI personas + 4G mobile modems — Week 1: 17 accounts, 10 burned in 48h, everything we learned

Weekly journey, real numbers only. What we did, what worked, what blew up. No fake screenshots, no "$10k in 3 days". Week 1 = mostly lessons, and some cost us accounts.

The goal
Build a network of aged, real-looking Reddit accounts to later:

• promote our own tools in relevant niches
• drive traffic to OnlyFans creators we work with

Right now we promote NOTHING. Phase 1 = pure account building: age, karma, clean history, zero removals. Promo on a 10-day / 12-karma account = burned account. That comes later, and I'll document the switch.

The stack

• Proxies - dedicated 4G mobile modems (PL carrier). 1 modem = 1 account, never shared.
• ReddMator - cloud panel for many accounts: warmup, comments, AI posts, engage, limits, schedules, per-account proxy.
• FluidTalk - AI persona engine. Each account = own persona (age, city, job, bio, voice, topics) + own API key. No two voices alike.

Timeline

• Day 1-2 (Sep 17-18): 1 account, warmup only (3 sessions/day). Day 2: AI comments 1/h, 3/day -> landed + upvotes, no shadowban.
• Day 3 (Sep 19): scaled to 17 accounts - 7 on 4G modems, 10 on "premium" residential (FR+DE, sticky 60 min). Own persona + 6 target subs each, no overlaps.
• Day 3-5 - the massacre: all 10 residential gone in ~48h. 4G: 6/7 alive.
• Day 6-7 (Sep 22-23): lost our best account (66 karma): flagged -> gone. Then read every AI comment (60+) - many sounded like AI. Big fix of the week.
• Day 8-10 (Sep 24-27): tuned persona engine, turned on replies + engage + AI posts (human approval queue), added new accounts on free modems - warmup only, no comments this time.

Proxy lesson: we chose residential because reddit.com returned 200 (403 on the modems), so it looked "cleaner". That signal was completely backwards. Don't judge a proxy by what reddit.com returns to an anonymous request.

Current settings (per account)


How the AI comments work
ReddMator finds a fresh post in a target sub -> sends it to FluidTalk with that account's persona key. The persona decides whether to comment at all. Skipping is normal - a thin comment is worse than none. If yes, ReddMator posts it within the account's limits and hours. Each persona is a full character (a 29yo physio from Leeds doing C25K, an HVAC tech, a night-shift biohacker...) with its own voice, slang and example lines.


Biggest lesson: reading 90+ AI comments

• Comment ending with a question to OP = 3x less karma (+0.27 vs +0.89 avg). Now banned.
• Generic filler ("consistency is key") came from the persona text, not the prompt. Persona beats prompt, every time.
• We over-corrected with an "every comment must add value" rule -> karma dropped to -0.25/comment (personas turned into know-it-alls). Fixed by splitting: question posts get a concrete answer, brag posts get a short warm reply with zero advice.
• Slang in every comment ("proper", "yeah nah") made accounts recognisable. Cut at the persona level.
• Subs with 3+ removals get paused and swapped - removals hurt the contributor quality score and they compound.

Numbers after 10 days

• 5 accounts commenting, ~113 comments, 162 comment karma total
• 3 of 5 commenting accounts have zero removals
• best single comment: +32 (a fitness sub)
• one account had 15 of 28 comments removed and sat at 0 karma for a week (wrong subs: loseit, TooAfraidToAsk, Fitness...). After two rounds of swapping to smaller niche subs it went 0 -> 50 karma in 3 days
• first AI posts approved and live (C25K, suggestmeabook, BuyItForLife)
• survival: residential 0/10 within 48h; 4G modems 5 of the original 7 still alive after 10 days (one flagged on day 3, our best account lost on day 6)

Plan for week 2

• get every account past 100 karma and ~30 days
• new batch: 5+ days of warmup before the first comment
• more 4G modems (we're out of free ones, and we're sticking to 1 modem = 1 account)
• measure AI posts: approval rate, removals, karma per post
• start planning the first soft promo tests on the strongest accounts

Questions welcome. I'll answer what I can without burning the accounts. Next update in a week.

Check PM

The proxy lesson was probably the most interesting part for me. Losing 10 accounts in 48 hours really shows how quickly things can go wrong when scaling too fast. Also the AI comment review was a good catch. Sometimes the small patterns in the writing are way more obvious than you think until you actually read a bunch of comments together.

Looking forward to learn something from this

Interesting. I'm an engineer building a very similar thing - my own custom version of these tools, and the ecosystem around them.

Reach out if you're interested, could be some useful things to share.

Tyziqjxor said:

is this your own tool this reddmator thing?

PS: interesting journey I will follow

Click to expand...

Thank you, I will write more about that software in future.

satyr85 said:

Check PM

Click to expand...

Replied.

MoMarketing said:

The proxy lesson was probably the most interesting part for me. Losing 10 accounts in 48 hours really shows how quickly things can go wrong when scaling too fast. Also the AI comment review was a good catch. Sometimes the small patterns in the writing are way more obvious than you think until you actually read a bunch of comments together.

Looking forward to learn something from this

Click to expand...

Thank you, however, accounts which used residential proxies was bought at accsmarket for 20 cents each, I think that was the case. Currently testing more accounts on residential which I have created manually, so far works fine. I will write more about that in next update.

the_ship said:

Interesting. I'm an engineer building a very similar thing - my own custom version of these tools, and the ecosystem around them.

Reach out if you're interested, could be some useful things to share.

Click to expand...

That's great, for sure would love to chat about process. Let's wait for my next update, for now seems everything is smooth.

In my batches, matching posting windows across accounts got them flagged faster than shared IPs ever did, so I'd compare the schedules of the 10 that burned.

shortline said:

In my batches, matching posting windows across accounts got them flagged faster than shared IPs ever did, so I'd compare the schedules of the 10 that burned.

Click to expand...

The thing is, that's not the case: the accounts weren't sharing the same IPs, and we split working hours by account.

Solid start. The insight about residential proxies being flagged faster than dedicated 4G is spot on for Reddit. Keep tweaking those AI personas to avoid the robotic tone, it makes a huge difference in longevity.

Interesting journey, I'm definitely following this one. The FluidTalk persona part caught my attention, especially how much the persona affects the comments. Curious to see how it goes once you start adding more accounts.

Now that’s how an experiment is done... break half the setup, find the flaw, and outsmart the algorithm on week two
Definitely curious about that AI persona engine you're running, though
How are you training its brain to blend in with the subs so smoothly?

Mobile proxies are truly the cleanest IP. Always trusted

Week 2 update (Sep 28 – Oct 8)
Same rules as week 1: real numbers, losses included.

TL;DR

• 1 account lost this week (week 1: 12) — the weak gaming account, suspended Oct 8
• fleet: 10 → 30 accounts (20 new = warmup only)
• fleet karma: 171 → 1,167
• commenting accounts: 5 → 10
• this week: 306 comments, 166 engage replies, 34 AI posts (587 upvotes, 358 comments)
• best comment +70, best post +187
• still promoting nothing

What we changed

AI posts = the surprise of the week

• 34 posts = 10% of what we published, 45% of the karma
• avg 17.3 upvotes/post vs 2.3 per comment
• wins: rants, hot takes, confessions in big taste subs (books, film, night shift, C25K)
• flops: questions in advice subs (~1 upvote)
• 6 posts never went live — sub rules (flair, trusted-only, 300-char limit, 2 sub bans)

Comments: the sub matters more than the words

• subreddit explains ~26% of a comment's score, account ~9%, writing style ~4%
• taste subs (A24, vinyl, Letterboxd, aquariums): 2.5–9.2 karma/comment
• advice subs (HVAC, appliance repair, "should I buy"): 0.4–0.7
• top 10% of comments = 45% of all karma
• length didn't matter. One sharp opinion > a 5-sentence lecture

Engage + replies under own posts

• engage = persona jumps into an existing conversation (replies to a user, not to OP)
• 166 engage replies this week, 0 removed
• 127 times the persona skipped (comment too short / nothing worth adding) — skipping is a feature
• 38 failed with 403 — mostly on 3 weaker accounts, kept at low volume
• new: replies under our own posts (max 3/post). Posts with 20–90 comments now get answers from OP
• no score data for engage yet — the panel doesn't sync it, so no karma claims here

@Trebu — "how are you training its brain?"
No training. It's depth, not a smarter model:

• week 1 personas: 4 facts, empty knowledge base → engine refused most posts ("thin fit")
• 525 scans in 7 days → 431 refusals → 323 of them were killed by our own quality floors, not by the engine
• fix: 12 facts per persona (job, home, pets, gear, history, opinions) + knowledge base + example lines
• result: 50 drafts in the queue in one day, quality score 85–100
• persona also decides to skip — most of the "blending in" is not posting when it has nothing to say

Proxy update (as promised to @MoMarketing)

• week-1 residential = $0.20 bought accounts. This round: accounts we made ourselves, PL residential, 1 IP per account
• 4 accounts since Oct 7, 4/4 alive, warmup only
• catch: sticky residential sessions die after ~24h → IP changes daily, sessions get re-created
• too early to call — verdict in week 3
• 4G: moving to 2 accounts per modem on opposite 12h shifts (00–12 / 12–24 UTC) to scale without buying modems. 8 accounts on the night shift now — verdict later

What went wrong

• OF persona: 26 comments removed, 19 of them in 2 subs (creator-advice + a music sub). Best engagement in the fleet, worst removal rate. Subs cut
• comment HTTP 500s: 63 failed attempts, almost all on 3 accounts
• gaming account suspended (Oct 8) — 17 karma, high 403/500 error rate all week. We slowed it down instead of stopping it. Lesson: errors are the warning sign
• 47% of comments get no traction past their own upvote
• my own "success rate" KPI was lying (~90% "posted ok" ≠ anyone cared). Now tracking avg score + % with no traction

Plan for week 3

• 20 new accounts: warmup until day ~20, then comments → engage → posts (personas already built)
• verdict on self-made PL residential accounts vs 4G
• more posts in taste subs, fewer advice subs
• first soft promo test on the 2 strongest accounts (200+ karma) — planning only, will document

Questions welcome. Next update in a week.

curdoc said:

Thank you, however, accounts which used residential proxies was bought at accsmarket for 20 cents each, I think that was the case. Currently testing more accounts on residential which I have created manually, so far works fine. I will write more about that in next update.

Click to expand...

Ahh that type of accs usually sucks because they are created in bulk quantity and have upvote manipulation on them no matter how hard you warm them up they have died for me as well

how you calculated those karma stats. Are you using actual account karma changes or individual comment scores? And how did you come up with those percentages?