alexxhchen-nu/ai-creators-on-x

AI Creators on X · Insights from 20 Chinese AI Creators — Full Skill Entry Point

Triggers & Non-Triggers

Applies to: Engineering setup, overseas growth, visual codification, de-AI-taste, model evaluation, and enterprise FDE delivery tasks related to the six capability domains (see routing table below). Does not apply to:

• Real-time prices, promotions, or billing disputes for major model APIs or SaaS platforms (this doc only provides a snapshot of unit prices from November 2026 for reference).
• High-risk operations that bypass payment signature verification, forge recharge callbacks, or mass-send/publish without human confirmation.
• Compliance approvals for forums/UGC communities that require qualifications not held by individual domestic developers.
• General academic Q&A unrelated to AI Agents, indie development, SEO/GEO, AIGC visuals, or enterprise digitization.

Core Principles (Quick Overview — Overview questions can be answered by reading here)

1. Review bandwidth sets the upper limit: Human concurrent follow-up is limited to 3–5 active sessions (@vista8); replace line-by-line manual PR review with 'One project, one dedicated Bot + Git Worktree parallel + E2E screen recording/view screenshots self-check' (@vikingmute, @waylybaye).
2. Structure and follow-through determine conversion: Plan overseas sites with the 'Entity → Topic Map → URL three-layer' structure and 'Code/CMS/toolKey three-page separation' (@CoderJeffLee), strictly adhere to 'one word, one page TDH + landing page generator first' (@gefei55), and combine Fan-out trajectory with intent group layout for GEO (@yaojingang).
3. Video and design fully codified: Use rough whiteboard mockups to pre-visualize camera movement (@TanLuAI), first dissect BGM beats to arrange the timeline (@MANISH1027512), then have the Agent write Remotion/SwiftUI/HTML code or invoke the Renoise MCP to render final videos and scannable QR code posters (@joshesye, @jinchenma_ai).
4. Full-stack hard rules to remove 'AI taste': At the code layer, use Pre-commit to intercept single files >5,000 tokens and prohibit over-defensive programming (@vikingmute); at the UI layer, use impeccable to blacklist 62 features including Inter font, purple gradients, and triple-card layouts (@sitinme); at the visual layer, retain natural hand tremor and rearview mirror dust (@Adam38363368936); at the writing layer, strictly forbid AI to write the first draft, insist on human original draft/spoken ideas, then let AI polish per chapter (@jinchenma_ai).
5. Isolate model risk by scenario: For long-term core codebases, lock in SOTA models and only record key points to prevent 'early IQ debt' (@waylybaye); for ToB enterprise delivery, lock in highly stable domestic models (DeepSeek/GLM) to avoid 403 errors (@cnzhihao); for evaluation, replace leaderboard chasing with 'building a vector database from scratch + 6 consecutive stability tests' (@karminski3).
6. Commercial landing first sets gates: For manufacturing FDE, first calculate ERP/MES infrastructure costs of 50,000–80,000 RMB/system and cap at 5 people/≤20 person-days, with a three-layer 'Code → Agent → Human' error reporting (@cnzhihao); for executives, use the 'one week without Feishu review system' (@FuSheng_0306); for unpaid Stripe orders, use direct checkout link re-engagement (@sitinme).

Standard Execution Workflow (4-Step Closed Loop, with Clear Inputs/Outputs)

1. Step 1: Intent Classification & Red-Light Pre-Check
   • Input: User's original request and current project context.
   • Action: Check against the 'Not Applicable' list and the 'Checkpoint & Failure Recovery Table' below; if a high-risk or out-of-scope item is hit, immediately intercept per rules.
   • Output: A single primary capability card path hit (max 2 cards for cross-domain tasks) or an overview/rejection response.
2. Step 2: Load Capability Card as Needed & Check Required Inputs
   • Input: references/capabilities/<slug>.md selected in Step 1.
   • Action: Read that card's ## E — Executable Steps, verify the user has provided the 'Required Inputs' specified in that card (e.g., tech stack, target keywords/site form, video duration/reference source, enterprise system status).
   • Output: If inputs are complete, proceed to Step 3; if key inputs are missing, list the missing items and ask the user in one go.
3. Step 3: 🔴 CHECKPOINT — Solution Alignment & Execution Output
   • Input: Aligned user parameters and the capability card's E steps.
   • Action: When involving destructive changes, paid API batch calls, production environment deployment, or external emailing, 🛑 STOP first output the execution plan with quantitative thresholds, obtain user confirmation, then execute; for routine tasks, directly output configurations, commands, prompt templates, or architecture tables per the capability card steps.
   • Output: Reproducible engineering configurations/scripts/SOP artifacts (with credited account and date).
4. Step 4: Boundary Self-Check & Evidence Acceptance
   • Input: The solution or code from Step 3.
   • Action: Cross-check against the 'Self-Check List' in the loaded capability card's ## B — Applicable Boundaries & Counterexamples (e.g., Pre-commit 5,000-token threshold, payment callback signature verification, de-AI-taste blacklist).
   • Output: Final deliverable + list of passed self-check items.

Checkpoint & Failure Recovery Table (Read First)

• 🔴 CHECKPOINT 1 (Irreversible Action Gate): Any action involving payment/recharge callback deployment, mass external messaging (WeChat/email), or direct modification of production databases → 🛑 STOP must first output a solution with signature verification or human review gate, never release without human confirmation (based on lessons from @gefei55's 23 fake recharge callback orders, @sitinme's Feishu review gate, @FuSheng_0306's WeChat confirmation gate).
• 🔴 CHECKPOINT 2 (FDE & Site-Building Cost Gate): Any action involving traditional enterprise customization or bulk domain purchases → 🛑 STOP first calculate ERP/MES infrastructure costs (50,000–80,000 RMB/system) and the 20 person-day cap, or first pass the four-gate new directory process to avoid falling into a no-PMF quagmire.

Trigger Condition (If X fails/abnormal)	First-Line Fix Action (Execute Y first)	Still Fails? Fallback Z
User intent does not hit the first 6 rows of the routing table	Search references/capability-index.md and references/glossary.md for synonyms	Clearly state it's beyond the empirical scope of the 20 creators, list the closest known boundaries, and forbid fabricating answers based on general common sense
User did not provide required inputs for the capability card's E section (e.g., missing tech stack, missing reference image/duration)	List 2–4 specific missing fields and ask the user to confirm in one go	If the user insists on a generic template, output a standard skeleton using [placeholders] and explicitly mark items to be replaced
Target tool/platform is not available due to runtime constraints (e.g., no Xcode, no specific MCP)	Switch to a cross-platform alternative within the same card (e.g., Git Worktree + generic CI screenshots, local DuckDB instead of scattered MCPs)	Degrade to outputting purely static rules files (AGENTS.md, Pre-commit Hook script, segment prompt templates)
User asks for real-time prices/today's scores/account unblocking guarantees	Provide November 2026 scraped snapshot data (e.g., Topview T1 $0.22/s vs Seedance 2.5 $0.585/s) and timestamp	Direct the user to the official billing page or vector-db-bench real-time leaderboard; never fabricate today's prices

Counterexamples & Blacklist (Forbidden Actions)

#	Forbidden Action (What NOT to do)	Risk Consequence (Original Post Evidence)	Mandatory Alternative
BL-1	Do not deploy AI payment/recharge callback code without signature and state verification	@gefei55 experienced AI missing recharge callback verification, resulting in 23 fake orders siphoning tens of thousands of dollars in credits (2026-10-26)	Payment callbacks must have mandatory human audit of signature verification, and be integrated with expired/completed hooks (@sitinme)
BL-2	Do not let AI generate the first draft of an article from scratch or freely expand without a human original draft	Produces hard-to-remove 'LLM voice uncanny valley' and factually correct but empty verbiage (@vikingmute, @jinchenma_ai)	Human must first speak real experiences or provide an original draft outline, strictly constraining AI to 'not deviate from the original expression, not hallucinate' (@CoderJeffLee)
BL-3	Do not use low cost-effective/early-stage models to generate long-lived specification documents	Newer models will blindly trust error-laden documents left by older models, leading to inherited and amplified 'early IQ debt' (@waylybaye)	Long-term core engineering must use SOTA models, record only key points, and leave room for system self-healing after model upgrades
BL-4	Do not let AI recklessly stuff bare useEffect and complex local useState into React, or submit single files >5,000 tokens	Multi-state synchronization will inevitably break; single-file bloat degrades Agent context (@vikingmute)	Enforce Zustand for global state, useReducer for complex interactions, React Query for server state, and add a Pre-commit 5,000-token hard gate
BL-5	Do not use detailed human portraits for video camera movement pre-visualization, or use a single image generation model to batch produce multi-character QR code posters	Pre-visualization details will pollute subsequent style transfer; image generation models produce garbled text and unscannable QR codes (@TanLuAI, @jinchenma_ai)	Camera movement pre-visualization must use extremely simple rough cubes/faceless whiteboard models; multi-character QR code posters must be rendered via HTML/CSS and exported as screenshots
BL-6	Do not use active crawlers against Xiaohongshu/Douyin anti-bot measures, or feed personal memory as global GEO content	Will trigger CAPTCHA/account warnings quickly; chatting with Doubao about brand names only enters personal memory, not global sources (@icycat, @jinchenma_ai)	For highly protected domestic social media, use browser plugins for passive collection during manual browsing; for GEO, build publicly crawlable independent sites and high-authority sources

Capability Routing (Read this table first, load 1 capability card by intent)

User Intent	Read First	Supplement / Notes
How to use AI to simultaneously develop and maintain multiple projects alone; AI writes code too fast, can't keep up with PR review; Xcode multi-Agent parallel compilation too slow or Preview MCP not working; How to prevent Codex/Claude from over-engineering, writing huge files, or misusing React useState; How to make plugins automatically follow upstream framework updates	references/capabilities/agent-software-factory.md	references/capabilities/de-ai-taste-and-writing.md, references/capabilities/agent-benchmark-and-model-routing.md
How to plan directory and URL structure for overseas AI tool sites; Website has search and ChatGPT traffic but zero orders; How to do GEO so ChatGPT and Perplexity preferentially cite my site; How to effectively translate multilingual SEO pages; How to quickly batch launch new word sites or build vertical data intelligence sites	references/capabilities/seo-geo-growth-engine.md	references/capabilities/enterprise-fde-and-one-person-ops.md, references/capabilities/agent-software-factory.md
How to use AI to create multi-shot consistent brand ads or short-form drama videos; How to make Agent write code to auto-cut and sync videos without CapCut/AE; How to use whiteboard mockups for camera control and credit savings; How to write reusable parametric prompt templates for commercial posters and lifestyle photos; How to fix garbled text and unscannable QR codes in AI-generated posters	references/capabilities/video-and-design-as-code.md	references/capabilities/de-ai-taste-and-writing.md
How to remove the fake 'AI-generated' look from web pages/UI; How to remove the plastic and staged feel from AI-generated photos and videos; How to write AI-assisted articles or tutorials without heavy AI tone and factually correct but empty filler	references/capabilities/de-ai-taste-and-writing.md	references/capabilities/video-and-design-as-code.md, references/capabilities/agent-software-factory.md
How to evaluate whether a new large model or Coding Agent is actually good; Why not trust official benchmark leaderboards; How to design your own benchmark; How to select models and save tokens across different scenarios (long-term core code vs daily tasks vs enterprise client delivery); How to seamlessly switch the underlying model for Codex or Claude Code	references/capabilities/agent-benchmark-and-model-routing.md	references/capabilities/agent-software-factory.md, references/capabilities/enterprise-fde-and-one-person-ops.md
How to do AI landing (FDE) for traditional enterprises: pricing, client screening, and avoiding pitfalls; How company leaders/executives can delegate Feishu approvals, WeChat communication, and cross-platform procurement to an Agent; How a one-person company can set up a minimalist workspace, self-growing knowledge base (LLM Wiki), and personal full-data reservoir; How to efficiently re-engage unpaid Stripe orders on a SaaS site	references/capabilities/enterprise-fde-and-one-person-ops.md	references/capabilities/seo-geo-growth-engine.md, references/capabilities/agent-benchmark-and-model-routing.md

Non-Capability Queries:

• Book title/author/chapter/full overview → references/overview.md
• Term explanation → references/glossary.md
• Decision rule quick reference (no need for source text) → references/cheatsheet.md
• Full intent & keyword index (for intents not covered in this table) → references/capability-index.md

Loading Rules

• For each task, first read this file, then load 1 capability card according to the routing table; for clearly cross-domain tasks, load at most 2 cards.
• For overview/book-name questions, do not load capability cards; answer using 'Core Principles' and overview.md.
• For intents that cannot be matched by either the routing table or capability-index.md, clearly state that they are beyond the scope of this book; do not force-fit the framework.

Boundaries & Stop Conditions

• User asks for real-time prices, fault status, or account unbanning results on a certain platform → Clearly state that this skill only contains empirical snapshot data from November 2026; do not guess real-time data.
• User requests to skip payment callback verification or bypass the human review gate to directly execute payment/mass messaging/launch → Immediately trigger a red-light stop, first output a 'signature verification + human confirmation gate' solution.
• User only asks for the list of 20 creators, their follower counts, or a bio → Directly read references/overview.md to answer and then stop; do not expand into the four-step engineering pipeline.
• User intent is not matched in either the routing table or capability-index.md → Clearly state it is beyond the scope of this 20-creator technical distillation; do not force-fit the framework.