Datasets:
Twitter/X Agent Operations — The Complete SOP for AI-Automated Account Management
Battle-tested: An AI agent grew @WeiYipei from 1,150 → 1,837 followers (+60%) in 45 days, posting 1 tweet per day, fully automated.
This skill works with any AI agent that supports a system prompt (Claude Code, Cursor, Trae, GPT).
1. System Architecture Overview
┌─────────────────────────────────────────────────────────┐
│ Twitter Agent Operations │
├─────────────────────────────────────────────────────────┤
│ │
│ [1] Persona ──→ [2] Source Library ──→ [3] Scheduling │
│ │ │ │ │
│ ▼ ▼ ▼ │
│ [4] Hard Rules [5] Pre-publish QC [6] Tracking │
│ │
│ ───────────── Weekly Loop ───────────── │
│ Weekly report → Review → Adjust weights → Next week │
│ │
└─────────────────────────────────────────────────────────┘
Six core modules:
- Persona calibration system — make the agent write content that "sounds like this person"
- Source library — a fact-source database, zero fabrication
- Scheduling system — weekly content calendar, 1 tweet/day cadence
- Hard rules — non-negotiable safety lines
- Pre-publish checks — triple-translation + 5-point QC
- Data tracking — tweet-log + weekly reports + follower tracking
2. Persona Calibration System (Voice Guide)
Why You Need It
The biggest problem with AI agents isn't that they "can't write" — it's that what they write "doesn't sound like this person." Persona calibration solves the soul problem, not the formatting problem.
Calibration Steps
Step 1: Collect Raw Voice Samples
Gather authentic expression samples from the account owner (at least 3 levels of "concentration"):
| Concentration | Source | What it gives you |
|---|---|---|
| Strong | Blog posts / long-form articles | Narrative structure, value expression |
| Medium | Original social media posts | Conversational tone, fragmented expression |
| Light | Podcasts / interviews / conversations | The most authentic voice, verbal tics |
Step 2: Extract Iron Rules
Distill 3-5 non-negotiable expression rules from the samples. For example:
Rule A: Openers must start with "I" + a specific experience/number/moment
Rule B: No three-part essay structure (claim → evidence → call to action)
Rule C: Not every tweet needs a "takeaway"
Rule D: Bold = expressing conviction, not highlighting key points
Step 3: Build a Dead-Opener Blacklist
Mine historical data for "opener patterns with the lowest impressions" and explicitly ban them:
❌ Opening with someone else's quote
❌ Grand-thesis openers ("A counterintuitive thing about the AI era")
❌ Subject-less buzzword openers
❌ Opening directly with a philosophical one-liner
Step 4: Define Content-Type Weights
Allocate type ratios based on data performance:
| Type | Recommended share | Why |
|---|---|---|
| Long-form (personal experience + data + insight) | 70% | Where the viral hits cluster |
| Tool / resource posts | 20% | Highest bookmark counts |
| Life fragments / rants | 10% | Keeps the account human |
Field data: After dropping the "short aphorism" type, weekly average impressions rose 266%.
3. Source Library (SOURCE-INDEX)
Core Principle
Never fabricate data. Every specific number in a tweet must have a real, traceable source.
Build Steps
Step 1: Collect First-Hand Material
Convert all of the account owner's first-hand content into searchable text:
- Podcasts / interviews → full transcripts (whisper / manual)
- Articles / documents → archived in markdown
- Talks / presentations → key-point extraction
Step 2: Build the SOURCE-INDEX
Annotate every key fact point:
| Fact point | Source | Original location | Usable |
|--------|------|----------|---------|
| 6,000 stars in the first week of open-sourcing | ep01, line 77 | "we hit 6,000 stars in the very first week" | ✅ |
| 643 investors | ep06, line 32 | "we must have added 643 of them, if I remember correctly" | ✅ |
Step 3: Verify Regularly
Every week, check that the data points cited in the schedule still match the originals. Numbers stated in different podcasts/settings may differ — pick the most reliable version and annotate it.
Lesson from the field: A user once flagged a mix-up between "6,000 stars in three days" and "6,000 stars in one week." After verification, all original sources consistently said "the first week."
4. Scheduling System
Cadence
- 1 tweet per day (hard rule, never exceed)
- Publish time: a fixed window (recommended 14:00-15:00 Beijing time, or whenever your target audience is most active)
Schedule Template
Generate next week's schedule every Sunday:
## Monday ｜ [Type] ｜ [Topic]
**Source**: [specific entry in SOURCE-INDEX]
**5-point self-check**: ✅/❌
**Triple-translation self-check**: ✅/❌
**CTA (in replies)**: [link]
Dedup Mechanism
Before scheduling each tweet, check the tweet-log:
- Has the same core argument been posted in the past 30 days?
- Has the same data point been used in the past 14 days?
- If duplicated → change the angle or change the topic
Time-Window Reference (based on 3,861-tweet dataset analysis)
| Window | Best-suited content |
|---|---|
| 10:00-13:00 | Tools, tutorials, resource entry points |
| 17:00-23:00 | Flagship content, opinions, case breakdowns |
| 00:00-01:00 | High-bookmark content, developer tools |
Best monthly ranking: 17:00 > 23:00 > 13:00 > 11:00 > 20:00
5. Hard Rules
Absolutely Non-Negotiable:
| # | Rule | Explanation |
|---|---|---|
| 1 | Never fabricate data | Every number must have a real source; if there's no source, don't write it |
| 2 | 1 tweet/day | No overposting. The agent must not adjust the frequency on its own |
| 3 | No CTA in the tweet body | External links go in the 1st reply (the X algorithm penalizes in-body links by 30-90%) |
| 4 | Data alignment | Dynamic numbers must be fetched fresh before publishing |
| 5 | Triple translation | Every tweet must pass the triple-translation check (see below) |
| 6 | 5-point check | Every tweet must pass the 5-point check (at least 4/5) |
6. Pre-Publish Checks
Check A: Triple Translation (from internal language to external language)
Read each tweet through once and confirm there is no "announcement-speak":
| # | Translation | Before | After |
|---|---|---|---|
| 1 | Launch → Help | "We shipped a new feature" | "This feature turns your 80-page report into a 3-page distillation" |
| 2 | Capability → Scenario | "Supports long context" | "Reads an entire industry report in one pass and spots competitor changes" |
| 3 | Conclusion → Evidence | "It works great" | Show real screenshots, inputs/outputs, steps, comparisons |
If any sentence in the tweet reads like "We launched X / We upgraded Y" → it must be rewritten.
Check B: The 5-Point Check (one strong tweet = one mini information product)
| # | Check item | The reader question it answers |
|---|---|---|
| 1 | A value promise that's clear at first glance | "What does this have to do with me?" |
| 2 | One concrete usage scenario | "When would I actually use this?" |
| 3 | A barrier-lowering step or entry point | "Can I start right now?" |
| 4 | Screenshots, numbers, cases = evidence | "Why should I believe you?" |
| 5 | One reason worth bookmarking or sharing | "Why would I keep this around?" |
Fewer than 4/5 = don't post. Go back and revise.
7. Data Tracking
Tweet Log (record every single tweet)
| Date | Time | Tweet ID | Type + summary | Impressions | Engagement | Notes |
Weekly Report Template
Generate weekly:
- Follower change (start/end + daily gain)
- Analysis of the Top 3 posts by impressions
- Content-type performance comparison
- Strategy adjustment suggestions for next week
Key Metrics
| Metric | Meaning | Optimization direction |
|---|---|---|
| Bookmarks | More important than likes (a trust signal) | Tool/resource posts are naturally bookmark-heavy |
| Impressions | Algorithm distribution effectiveness | The opener decides 80% |
| Engagement rate | Content resonance | Comments > likes > retweets |
| Daily follower gain | Growth health | Steady > volatile |
8. Content Methodology Reference
Four Content Archetypes (based on 3,861 tweets)
| Type | Probability of reaching top 10% | Characteristics |
|---|---|---|
| Resource gateway | ~51% | Finds the entry point for readers (saves searching) |
| Tool tutorial | ~39% | Explains complex things for readers (saves understanding) |
| AI-tool discovery | ~24% | Shows a new tool + a concrete task (saves trial-and-error) |
| Plain opinion | ~9% | Pure opinion with no action (avoid) |
The Four-Savings Model
The value of content isn't how much information you say — it's how many steps you save the reader:
- Save searching — readers don't have to hunt for the entry point in an ocean of information
- Save understanding — readers don't have to puzzle out complex concepts themselves
- Save trial-and-error — readers don't have to step on every landmine themselves
- Save expression — readers can forward your tweet to someone else as-is
Three Visibility Principles
Readers are more willing to trust content that is visible, clickable, and quantifiable:
| Principle | Examples | Probability of reaching top 10% |
|---|---|---|
| Visible | Screenshots, screen recordings, comparison charts | — |
| Clickable | Links, tool names, search paths | ~40% (with "link in replies") |
| Countable | Numbers, time, cost, steps | ~35% (with resource keywords) |
Golden Length
| Characters per post | Probability of reaching top 10% |
|---|---|
| ≤40 chars | ~7% |
| 41-100 chars | ~15% |
| 120-220 chars | ~26-28% (golden zone) |
Template: First sentence states the value → sentences two and three describe the scenario → then give evidence or steps → end with an entry point or a reason to bookmark.
9. Case Study: @WeiYipei Operation Data
Growth Curve
Week 1 (4/24-4/28): 1,150 → 1,155 (+5) ← Cold start, exploration phase
Week 2 (4/28-5/05): 1,155 → 1,180 (+25) ← Started daily long-form posting
Week 3 (5/05-5/12): 1,180 → 1,250 (+70) ← First viral hit appeared
Week 4 (5/12-5/18): 1,250 → 1,380 (+130) ← Threads + engagement strategy
Week 5 (5/18-6/01): 1,380 → 1,540 (+160) ← Steady long-form output
Week 6 (6/01-6/08): 1,540 → 1,837 (+297) ← 40-Playbook panorama breakout
Total: 1,150 → 1,837 = +687 followers (+60%) in 45 days
Key Turning Points
| Event | Impact |
|---|---|
| Dropped the "short aphorism" type | Weekly impressions +266% |
| Fixed 8 AM publishing | Viral hit rate from 5% → 15% |
| Openers must be "I" + specific experience | All 6 viral hits were first-person |
| Thread (7-8 posts) power move | A single Thread gained 50-100 followers |
| 40-Playbook panorama | +297 followers in one week |
What Works vs. What Doesn't
| ✅ Works | ❌ Doesn't |
|---|---|
| Long-form + real experience + data | Philosophical one-liners / quoting others |
| Tool posts + weekend 8 AM | Posting in the dead of night (impressions <200) |
| CTA in the replies | CTA in the body (cut by 30-90%) |
| First-person openers | Grand-thesis / preachy openers |
| Steady 1/day cadence | 3 tweets in one day or 3-day gaps |
10. Quick-Start Guide
If You Want to Use This SOP Right Now:
Day 0 (Preparation, 2-3 hours):
- Collect 10 representative pieces from the account owner
- Extract 3-5 persona iron rules
- Build the dead-opener blacklist
- Set content-type weights
Day 1 (Source library, 2-4 hours):
- Convert all first-hand content into text
- Build the SOURCE-INDEX (key data points + provenance)
- Mark which points are usable and which need verification
Day 2 (Scheduling + rules, 1 hour):
- Write the first week's schedule (7 tweets)
- Confirm the hard rules
- Set the publishing time
Day 3 onward (Execution):
- Draft according to the schedule every day
- Run triple-translation + 5-point checks before publishing
- Record the tweet-log after publishing
- Produce a weekly report + adjustments every week
11. Common Mistakes
| Mistake | Consequence | Fix |
|---|---|---|
| Agent fabricates data | Trust collapses once users/the owner finds out | Hard rule 1 + mandatory SOURCE-INDEX |
| Overposting (multiple per day) | Algorithm downranking + content dilution | Hard rule 2, hard limit |
| Every tweet reads like an announcement | Impressions <300 | Triple-translation check |
| Aphorisms / preachy tone | Impressions 100-250 | Dead-opener blacklist |
| No data tracking | Impossible to optimize | Weekly report mechanism |
| Voice drift | Followers feel "this doesn't sound like them anymore" | Re-calibrate against the voice samples monthly |
Install
# ClawHub
clawhub install gingiris-twitter-agent-ops
# skills.sh
npx -y skills add Gingiris-1031/gingiris-twitter-agent-ops
# Or just copy this file into your AI agent project
Related links:
- HuggingFace: https://huggingface.co/datasets/Gingiris/gingiris-twitter-agent-ops
- GitHub: https://github.com/Gingiris-1031/gingiris-twitter-agent-ops
- More playbooks: https://gingiris.tools
Credits
- Methodology foundation: Xiangyang Qiaomu, "X Growth Experience: From 100 to 110K Followers" (analysis of 3,861 tweets)
- Content diagnosis framework: dontbesilent/dbskill "Content Creation Diagnosis"
- Field validation: the @WeiYipei account (operated by a Cola AI agent, April-June 2026)
- Author: Iris Wei (生姜iris) | Twitter @WeiYipei | https://gingiris.tools
License: MIT