Turn any AI agent into a ghostwriter that sounds like you — and grew a real account +60% in 45 days.
By Iris Wei (生姜) · ex-COO of AFFiNE (60K+ GitHub stars) · 30× Product Hunt #1
You know you should post daily, but writing eats your evenings — and when AI writes for you, it sounds nothing like you. This SOP turns any agent into a ghostwriter that actually sounds like you, with a fact-source database so it never fabricates numbers, and safety rails so it never embarrasses you. Real result: +60% in 45 days.
| Follower growth in 45 days | +60% |
| Posting cadence | 1 tweet/day |
| Tweet dataset analyzed | 3,861 tweets |
| Hard safety rules | 6 |
1 · Calibrate the persona
Build a voice guide and expression rules, plus a dead-opener blacklist, so the agent writes in your voice instead of generic AI cadence.
2 · Build the fact-source database
Create a SOURCE-INDEX of traceable facts and numbers. The agent may only cite stats that exist in the index — this is how it never fabricates.
3 · Schedule the week
Plan posts on a weekly schedule with dedup checks against past tweets and golden time windows (derived from a 3,861-tweet analysis).
4 · Pre-publish QC
Run every draft through the six hard safety rules and a pre-publish checklist, including a triple-translation test, before it goes live.
5 · Analytics loop
Keep a tweet-log and weekly report so the agent learns which archetypes and lengths perform, and feeds that back into scheduling.
The goal isn't to remove you — it's to keep your voice while reclaiming your evenings.
| Dimension | Manual | Agent-operated (this SOP) |
|---|---|---|
| Time cost | Hours/week | Minutes of review |
| Voice | Yours | Yours (persona-calibrated) |
| Fabricated stats | Human error | Blocked by SOURCE-INDEX |
| Consistency | Drops when busy | Daily, scheduled |
Can an AI agent really post in my voice?
Yes, with persona calibration — a voice guide, expression rules, and a dead-opener blacklist — validated by a triple-translation test before publishing. The case study grew a real account +60% in 45 days, fully agent-operated, without sounding like a bot.
How does it avoid making up statistics?
A fact-source database (SOURCE-INDEX) holds every traceable fact. The agent is constrained to cite only numbers that exist in the index, so it can't fabricate.
Is it safe to let an agent post automatically?
Only with rails: six hard safety rules plus a pre-publish QC checklist (including a triple-translation test) gate every post. The point is consistency with your voice, not unattended risk.
Who built this?
Iris Wei (生姜) — ex-COO of AFFiNE (60K+ GitHub stars), from analysis of a 3,861-tweet dataset and a real +60%-in-45-days account.