Ai Industry Infographic

Generates verified, ChatGPT-ready prompt packages for AI industry infographics, including topic design, data research, verification, prompts, and social medi...

AI Industry Infographic Generator

Generate complete, ChatGPT-ready infographic prompt packages for AI industry topics. Each package includes verified data, structured prompts, and social media copy.

Workflow

Step 1: Topic Design

Classify the infographic into one of these visual types based on data shape:

• Multi-layer architecture → Vertical stack diagram (e.g., industry chain layers, tech stack)
• Time-series comparison → Timeline + funnel chart (e.g., market consolidation, year-over-year evolution)
• Competitive comparison → Radar chart + bar chart (e.g., China vs US, multi-company benchmarks)
• Cost breakdown → Pie/donut chart + waterfall (e.g., BOM analysis, price decline curves)
• Ecosystem map → Radial/constellation diagram (e.g., corporate structure, investment network)
• Scenario matrix → 2×2 matrix + progress bars (e.g., deployment maturity by sector)

Step 2: Parallel Web Search

Execute 3-5 simultaneous searches from different angles. Typical angles:

1. Official/authoritative data (government reports, official exhibitor lists, financial filings)
2. Industry analyst coverage (甲子光年, 36Kr, 凤凰科技, 界面新闻)
3. International comparisons (English-language sources for global benchmarks)
4. Niche/supplementary data (specific company financials, supplier lists, historical archives)
5. Recent news (past 48 hours for latest developments)

Search in both Chinese and English. Prioritize primary sources over secondary reporting.

Step 2.5: Red Flag Pre-Screen (BEFORE verification)

Purpose: Before spending 3 rounds verifying data, immediately identify and remove data patterns that have a near-100% failure rate in prior verification runs. This saves 2-3 rounds of unnecessary cross-verification on data that will never pass.

Run this 8-item checklist on every search result before proceeding to Step 3:

#	Red Flag	Action
1	"XX是YY的独家供应商"	Flag 🔴. Remove "独家". If the supplier relationship is real, keep company name without "独家".
2	进口高端产品 vs 国产低端产品价格对比（如"进口12万 vs 国产3000元"）	Flag 🔴. Delete the comparison. Different specs/tiers cannot be directly compared.
3	"全球第一""全国首家"无具体统计范围和来源	Flag 🔴. Delete superlative unless S-grade source specifies exact scope and methodology.
4	精确到个位数的异常高数字（118亿/17亿）	Flag 🔴. Return for source verification. If no S-grade source, delete the number.
5	BOM成本百分比未指定具体产品和测算机构	Flag 🔴. Downgrade to 🟡 B-level and annotate: "XX方案下XX机构测算".
6	"订单排到20NN年"	Flag 🔴. Delete. Replace with qualitative milestone from company announcement.
7	国产化率精确到个位（如73.5%）	Flag 🔴. Convert to range or qualitative description. No unified statistical authority exists.
8	"估值/收入比XX倍""市场规模仅XX亿"	Flag 🔴. Annotate source. If no S/A-grade source, delete or use qualitative language.

After pre-screen: Only data that PASSES all 8 checks enters Step 3 verification. Flagged data is either deleted, or annotated as 🔴 and excluded from any output file. This is NOT a suggestion — it's a hard gate.

---

Step 3: Data Verification (CRITICAL — 3-pass protocol)

Single-source reporting frequently contains inaccuracies. Every key data point must survive a 3-pass verification before entering the output.

Pass 1: Source Grading

Classify each source before using its data:

Grade	Examples	Reliability	Rule
S	公司官网公告、招股书、监管文件、央媒首发（新华社/中新社）、政府白皮书	Highest	Can be primary source
A	证券时报/中国证券报/上海证券报/第一财经、IDC/高工等权威第三方数据	High	Must cross-verify with at least one S or another A
B	36Kr/虎嗅/界面/凤凰科技/甲子光年/每日经济新闻	Medium	Must cross-verify with at least one S or A
C	自媒体号/头条文章/公众号/小红书/微博	Low	Never use as sole source; S or A must independently confirm

Pass 2: Cross-Verification Rules

For each data point, apply these rules in order:

1. Business metrics (出货量, 融资额, 估值, 部署台数, 价格) → Require S + A or A + A. Never use a single B source alone.
2. Official announcements (首发, 镇馆之宝, 战略合作) → Require S or two independent A sources.
3. Narrative claims ("零失误", "全球第一", "行业首次") → Require S. If only B or C, downgrade to "据X报道" with explicit attribution.
4. Temporal data (成立时间, 发布日期) → Require S or A. Wikipedia/企查查 acceptable as A.

Pass 3: Conflict Resolution

When sources disagree on the same data point:

1. Prefer S over A over B over C — company's own announcement beats media paraphrasing
2. Prefer newer over older — later reporting often corrects earlier errors
3. Flag the conflict — if both sources are credible but disagree, use the higher-grade source and add a footnote: "注：另有报道称[不同数据]，本文以[来源]为准"
4. Example from WAIC project: 中国证券报(B) reported 智元在富临精工 "工位1→4, 负载5→14kg", but 智元官网(S) reported "作业点位2→15个". 智元官网 data used because S > B.

Red Flags That Trigger Re-Verification

• Any superlative claim ("全球首次", "唯一", "最大") without S-grade sourcing
• Exact round numbers (100, 1000, 10000) — often rounded or estimated
• Data attributed to "业内人士" or "据透露" without named sources
• Numbers that differ significantly between two B-grade media sources
• Statistics that would require access to non-public data (e.g., exact competitor shipment numbers)
• "独家供应商" claims — almost never confirmed in public filings. Downgrade to "已进入供应链" or softer language.
• Cross-tier price comparisons — comparing high-end imports to low-end local products is apples-to-oranges. Describe price ranges separately, don't use "vs".
• BOM claims without product/model specification — "执行器占BOM 30-60%" is meaningless without specifying which robot and which research institution.
• "订单排到202N年" — companies rarely publicly confirm multi-year forward order books. Use qualitative milestones instead ("已进入小批量生产").
• 国产化率 as exact percentage — no unified statistical authority exists for most niche component categories. Use ranges or qualitative descriptions.

⛔ Step 3.5: Mandatory 3-Round Verification (STRUCTURAL GATE)

通用版: 本协议已独立为 web-verify-protocol skill（三轮联网搜索验证协议）， 适用于任何"AI 出数据"场景。此处保留完整版供信息图上下文使用，修改时需两边同步。

Core principle: A single search pass catches ~70% of errors. A second catches ~90%. Only three independent passes catch close to 100%. This is not a suggestion or a "review checklist"—it is a structural requirement built into the workflow.

How this gate works: You cannot "interpret" your way around it. Below are hard quantitative minimums. If they are not met, Step 4 is blocked.

---

Round 1 — Fresh search, S-grade upgrade (min 3 WebSearch calls)

Search from angles that did NOT appear in Step 2. The goal is not to review what you already found—it is to find better sources.

Hard minimums:

• At least 3 separate WebSearch calls, each from a different angle
• At least 1 search must target S-grade sources specifically (company official announcements, Xinhua/CNS, government white papers, financial filings, gov.cn)
• Result: Every data point classified as S or A in Pass 1 (Step 3). Any data stuck at B-grade after R1 must be flagged as ⚠️ unverified.

What R1 catches: Media paraphrasing errors (e.g., "双足人形" vs "轮式"), imprecise metrics, wrong company attribution, outdated pre-event data used as current.

---

Round 2 — Independent cross-verification (min 3 WebSearch calls)

Take R1's verified data and search for a second independent source that confirms or challenges each key metric. You are looking for sources that did NOT appear in R1 results—new outlets, new angles, new search queries.

Hard minimums:

• At least 3 separate WebSearch calls
• At least 1 search must target factual disputes (search for "[claim] + 核实" or "[claim] + correction")
• Result: Every business metric confirmed by ≥2 independent S or A sources. Single-source metrics must be explicitly annotated: "单源确认，待交叉验证"

What R2 catches: Single-source errors propagated when media quote each other, outdated numbers republished as current, regional data misattributed as national.

---

Round 3 — Narrative claim audit (min 2 WebSearch calls)

This round ONLY targets storytelling numbers: percentage improvements, "X-fold" claims, before/after comparisons, superlatives, any number that feels "too neat."

Hard minimums:

• At least 2 separate WebSearch calls
• Each call must target a specific narrative number from your data tables (not a general topic search)
• Result: Every narrative claim must have an S or A source. Claims from B/C sources are either replaced with S-grade alternatives or explicitly attributed ("据XX白皮书") and marked ⚠️.

What R3 catches: B-grade white papers and analyst reports using imprecise storytelling numbers contradicted by S-grade company/government data.

---

⛔ BLOCKING GATE — Print this before ANY output file

You MUST print the following report in your conversation response BEFORE the first Bash call that writes a file. This is the user's only way to verify rounds were completed without asking "did you run three rounds?"

text

══════════ 三轮核实报告 ══════════
R1 S级升级: {N}次搜索, {M}个数据点从B升到S/A, {K}个仍为⚠️单源
R2 交叉验证: {N}次搜索, {M}个指标≥2源确认, {K}处修正
R3 叙事审查: {N}次搜索, {M}个叙事数字审查, {K}个替换/降级
结论: ✅ 通过 / ❌ 需补充{R1/R2/R3}
═══════════════════════════════

If any round has fewer than the minimum calls, or if R3 found unresolved narrative claims, the gate is NOT passed. Fix the gaps and re-verify.

After the gate: Proceed to Step 4 and write output files. The files you deliver are the final verified version—not a draft awaiting the user to ask for verification rounds.

---

Step 3.6: Data Confidence Levels for Infographic Output

After the 3-round verification gate, every data point must receive a confidence level that determines whether it can appear on an infographic, and with what annotation. This is separate from source grading (S/A/B/C for the origin) — it represents a final quality gate before visual output.

🟢 A-Level — Can appear on infographic without annotation

Condition: S-grade source (company filings/government docs) OR A-grade source with ≥2 independent confirmations.

Examples:

• SK海力士Q2毛利率83%（财报原文）
• 2026年人形机器人整机产量有望突破10万台（工信部）
• SEMI半导体材料国产化率15-20%（官方报告）

🟡 B-Level — Can appear on infographic WITH source annotation

Condition: B-grade source with ≥2 independent confirmations, OR A-grade from a single source. Must NOT contain superlative claims ("独家/第一/唯一/最大") without S-grade confirmation.

Annotation format: "据中国信通院2026Q1，多源引述" / "摩根士丹利2026研报测算" / "行业估计"

Examples:

• 行星滚柱丝杠国产化率约40-50%（据信通院Q1，多源引述）
• Optimus整机BOM 1.2-1.5万美元（摩根士丹利2026研报测算）

🔴 C-Level — CANNOT appear on infographic

Condition: Single B-grade source, OR contains unverifiable superlatives, OR the number is abnormally high/low without official confirmation, OR different sources for the same metric diverge by >30%.

Treatment: Either re-verify with stronger sources, or delete entirely. If the insight is true but the number can't be verified, use qualitative language instead.

Examples of what gets deleted:

• "绿的谐波市占60%+、订单排到2027年" → 未找到公告原文，删除后将"订单排到2027年"改为"已进入小批量生产"
• "进口丝杠8-12万/套 vs 国产3000元" → 量程/精度不同不可比价，删除价格对比
• "拓普Optimus独家供应商" → "独家"无官方来源，改为"布局并推进机器人电驱执行器产业化"

⚠️ 产品线差异化标准

Different output channels have different confidence thresholds:

产品线	🟢 A级	🟡 B级	🔴 C级
商业图解（信息图/付费图）	必须	允许，标注来源	严格禁止
公众号文章（ECS pipeline）	必须	允许，可弱标注	禁止；但可用定性描述替代
公众号素材（data pack for rewriting）	必须	允许	禁止出现在素材文件中

商业图解的信息图要求最严格——数字直接印在图上，不可撤回。公众号文章经deepseek改写后有缓冲层。

Pattern Library: Common C-Level Red Flags

The following patterns consistently fail verification and should trigger immediate re-verification before appearing in any output:

Pattern	Why it fails	Correct handling
"XX公司是YY的独家供应商"	"独家"几乎从不出现在公开公告中	改为"供应商"或"已进入供应链"
进口高端 vs 国产低端价格对比	规格/量程/精度不同，不可跨等级比价	分别描述价格区间，不写"vs"
"全球第一""全国首家"无范围限定	下线量≠交付量≠销量，统计范围不明	标注具体口径或删除"第一"
精确到个位数的订单/收入数字（118亿/17亿）	异常高数字通常来自单一B级源	退回核实公告原文，找不到就删
BOM成本外推到所有产品	Tesla Optimus的BOM≠宇树G1≠所有人形机器人	必须标注具体产品和测算机构
"订单排到202N年"	极少有公司公开确认远期订单	改为"已进入XX阶段"等定性描述
国产化率精确到个位（如73.5%）	这类数据通常不存在统一统计口径	用区间（40-50%）或定性（"仍较低"）
"估值/收入比XX倍""市场规模XX亿"	通常来自单一券商或媒体估算	标注"据XX测算"，或改为"据行业观察"

Correction Propagation Rule

When a data point is revised (downgraded or deleted) in one output, all other deliverables in the same project sharing that data must also be updated. This applies to:

• 图解资料包 → 同步修正对应的公众号素材文件
• 公众号素材 → 同步修正ECS inbox中已上传的素材（需重新scp）
• 交叉矩阵 → 如果数据出现在矩阵中，同步修正

Step 4: Structured Output

Write the output in this exact four-section format:

text

# 生图 NXX：{Chinese Title}

## 一、选题定位
- Title (one line)
- Hook (one sentence summarizing the key insight)

## 二、核心数据
- Structured tables with verified figures
- Comparison tables, timelines, or hierarchical breakdowns
- Annotate data sources inline

## 三、ChatGPT 生图 Prompt
- English prompt with clear layout instructions
- Specify chart types, color palette, data points to include
- Add: "No realistic faces or product photos. Dark tech background (#1a1a2e)."

## 四、发帖文案
- 120-200 characters
- One sentence per paragraph (一句一段)
- No markdown bold, no hashtags in body, no parallelism patterns
- Conversational, first-person tone with personal observations
- Close with a judgment, not a scenic ending

## 数据来源
- List all sources with dates

Step 5: Title + Copy Variations

Generate 5 title options for 小红书, ranked from data-shock to insight-driven:

1. Data shock (big number, surprising stat)
2. Practical utility (search-friendly,收藏-oriented)
3. Scarcity/unique angle (previously unpublished)
4. Identity hook (where do YOU fit?)
5. Knowledge summary (authoritative conclusion)

Generate 1 copy body following the tone rules above.

Design Rules

• Every data point must be attributable to a specific source
• Never fabricate exhibitor names — cross-verify against official lists
• When official lists are available (e.g., venue brand lists), use them as ground truth
• Mark unverified claims explicitly with "需核实" or "待确认"
• Store search data even for rejected angles — they become reference for future topics

Step 6: Post-Generation Correction Protocol

When the user or subsequent fact-checking discovers errors in a generated package:

1. Re-verify the disputed data point using the 3-pass protocol (Step 3)
2. Correct in-place — edit the source file directly, do not create a new file
3. Propagate corrections — if the same data appears in other packages (e.g., a company's shipment number used in N04, N08, N09), fix all occurrences
4. Add a correction note at the bottom of the file: 🔄 YYYY-MM-DD 修正：[具体修正内容]。来源：[S/A/B级来源]
5. Update the开工文档 if the data also appears in the master reference document

Example: 灵心巧手出货量从"月千台"修正为"月超4000台" → corrected in N12, propagated to any future N0X that cites this number.

Advanced Protocols

A. Topic Prioritization

Before generating a topic, classify its lifecycle:

Lifecycle	Examples	Rule
Event-driven (strong deadline)	WAIC闭幕数据、展会热点	Must publish within 48h of event; skip if missed
Evergreen (no deadline)	产业链图谱、技术栈拆解、商业模式分析	Can publish anytime; deprioritize during event crunch
Hybrid (event hook + evergreen body)	"WAIC上看到的具身智能落地信号"	Event hook has 48h window; body is evergreen. Publish within window.

Assign P0/P1/P2 based on lifecycle + data readiness:

• P0: Evergreen with all S/A-grade data verified OR event-driven within window
• P1: Evergreen with some B-grade data pending verification
• P2: Event-driven but window passed → demote or skip

B. Data Reuse Across Topics

When multiple topics share the same data (common in a series about one event):

1. Establish a single source of truth — put shared data in the开工文档 appendix, not duplicated in each topic file
2. Cross-reference by topic ID — in each topic file, cite "数据详见开工文档附录 N04" rather than re-copying
3. When data is updated — update in the appendix first, then check all referencing topics
4. Mark shared data as such: "📎 共享数据 — 与N04/N08共用，修改时需同步"

C. 小红书 Copy Tone Standards

After extensive iteration, the following rules produce natural, human-sounding copy:

Must do:

• Start with a personal observation or specific anecdote
• Use conversational connectors like "说实话", "有意思的是", "但真正让我意外的是"
• Include at least one concrete scene or moment (a number you noticed, a booth you walked past)
• End with a judgment, insight, or open question — not a summary

Must NOT do:

• No structural parallelism ("一方面...另一方面", "首先...其次...最后")
• No hashtags in body text
• No "在这个时代" "随着AI的发展" style of grand opening
• No exclamation marks as emphasis tools
• No explicit call-to-action ("关注我", "收藏这张图") — let the content earn the save

Title formula:

1. Data shock: "{Surprising number or stat}" — e.g., "三年降了300倍"
2. Utility: "{What problem this solves}" — e.g., "整理了两天 把中国AI产业链画成了这张图"
3. Scarcity: "{Previously unpublished angle}" — e.g., "还没人做过"
4. Identity: "{Where YOU fit}" — e.g., "如果你也在AI行业"
5. Knowledge: "{Authoritative conclusion}" — e.g., "从芯片到机器人 五层 缺一层都跑不起来"

D. Prompt Template Library

For each visual type, use this prompt skeleton:

2×2 Matrix:

text

Create a [title]. LAYOUT: 2×2 matrix. X-axis: [dim1]. Y-axis: [dim2].
Four quadrants: [TL/TR/BL/BR descriptions with data]. Below matrix: [summary stats].
DESIGN: Dark tech background (#1a1a2e). [Color scheme]. No realistic faces/photos.

Comparison Table:

text

Create a [title]. LAYOUT: Clean comparison table. Columns: [list]. Rows: [list].
Use color gradient: [scheme]. Below: [additional context].
DESIGN: Dark tech background (#1a1a2e). [Specific layout notes]. No realistic faces/photos.

Timeline + Funnel:

text

Create a [title]. LAYOUT: Horizontal timeline [years] above, funnel chart below.
Timeline annotations: [key events per year]. Funnel stages: [stages with numbers].
DESIGN: Dark tech background (#1a1a2e). Gradient from [color1] to [color2]. No realistic faces/photos.

Multi-Layer Architecture:

text

Create a [title]. LAYOUT: Vertical stack of [N] layers. Each layer shows: [name, key players, stat].
Right sidebar: [cross-layer trends]. Bottom: [summary].
DESIGN: Dark tech background (#1a1a2e). Color gradient bottom-to-top. No realistic faces/photos.

E. Collaboration with exhibitor-list-generator

When both skills are used in the same project:

1. Company data in info graphics — pull from the verified .xlsx, never from memory or B-grade media
2. Sector classification — use the same tag system from exhibitor-list-generator for consistency
3. When adding a company to an infographic — verify it exists in the verified list first
4. If the infographic needs a company not in the list — flag it for exhibitor-list-generator to verify and add

F. Obsolescence Rules

After a major event concludes:

1. Pre-event data (会前预测、展商预告) → Mark as "会前数据，以实际为准" or discard
2. Event-day data (首日直击、排队情况) → Retain as historical record, mark date prominently
3. Post-event analysis (闭幕数据、行业总结) → This is the canonical reference going forward
4. When the same metric appears in pre-event and post-event sources → Use post-event data (e.g., 实际参展数 > 预告参展数)
5. After 30 days → Review all topic packages for stale data; update or annotate as needed