Bottom line first: Channels like The Operations Room are a secondary processing layer on top of primary sources, not primary sources themselves. The real information-source hierarchy has three layers: archives/official documents → English explainer channels (processing layer) → Chinese repost accounts (terminal layer). Your opportunity lies in the gap between Layer 2 and Layer 3: build an automated pipeline that delivers “Layer 2 speed + Layer 3 language.” Map recreation is viable and constitutes a differentiation moat; English-to-Chinese copy translation plus LLM rewriting sits in a grey zone — doable but with a ceiling and compliance risk. Genre recommendation: expedition accident case studies > classic battle histories > disaster investigation reports. Avoid hot military news during the cold-start phase.

Lingering Questions from the Previous Article

After I broke down that Douyin military commentary video (frame-by-frame comparison confirmed it was reposted from The Operations Room, confidence 0.86), readers asked a string of genuinely actionable questions: Do these YouTube channels count as primary sources? Is it feasible to use a Windmill workflow plus Claude Code to orchestrate Python scripts for automated monitoring? Can you dodge plagiarism by directly adapting the video logic into “map recreation + English-to-Chinese copy translation with LLM rewriting”? What genres should you pursue?

This article is the answer. First I’ll unpack the concept of “primary source” cleanly, then lay out the pipeline design, and finally address topic selection and feasibility.

Information Sources Are Three Layers Deep — “Primary” Is a Misconception

The raw material behind the Python Expedition video actually looks like this:

The Operations Room itself doesn’t start from L0 — it stands on top of declassified archives and dozens of secondary studies to produce its visualizations. But what it does (transforming archives into animated map reconstructions with timelines) is irreplaceable processing, and that’s where the “primary source feel” comes from: what the audience perceives as “primary” is actually processing depth, not information-source level.

This means two things for your strategy: First, you don’t need to — and shouldn’t — jump to L0 to do archive research (that’s what channel owners spend hundreds of hours on per video). Second, the L1→L2 gap is a real, underserved market: the English-language world produces massive volumes of high-quality L1 content, while Chinese L2 supply is scarce and low-quality (repost accounts don’t even bother remaking the maps, they just screenshot them). That gap is your arbitrage space.

The Automated Pipeline: How the Three Components Divide Labor

Your tool stack happens to be an ideal configuration for three-layer orchestration:

1

Windmill (scheduling + triggers + state) → Hermes (agent orchestration + notifications) → Claude Code (code implementation + video scripting)

Windmill handles scheduling and state. You already have a CE instance running; add a monitor script: use yt-dlp to pull recent video metadata (title / publish time / duration / view count / growth curve) from L1 channels (The Operations Room, Kings and Generals, Battle Order, plus top channels in the expedition-accident niche), store it in a database, and diff for new items. Monitoring frequency of once or twice per day is enough — L1 channels update on a weekly or monthly cadence, not hourly. For trigger conditions, I recommend using “view-count growth slope” rather than “newly published”: only videos that enter a channel’s P25 view-percentile band 48–72 hours after publication make it into the candidate pool. This directly filters out mediocre source material.

Hermes handles intelligent screening and notifications. The diff output from Windmill gets pushed to Hermes, where an agent evaluates content: whether the genre falls in the safe zone (avoiding active conflicts), whether the narrative structure is complete (timeline + geography + life-and-death suspense as the three要素), and whether the Chinese-language space has already covered it (keyword search on the title). Passed items enter a “candidate pool” topic notification for you to decide whether to produce. Don’t fully automate this step — picking the wrong genre wastes everything downstream.

Claude Code handles production. Once you greenlight a topic, the production pipeline kicks off. This is where Python-script orchestration comes in (I’m giving architecture only here; implementation comes post-kickoff):

1. Source extraction: yt-dlp downloads the original video → ffmpeg extracts frames at 1fps → a vision model tags each frame (classifying map frames / person frames / archival footage frames) → outputs structured JSON. This is the three-channel extraction logic of video2script.
2. Copy rewriting: ASR transcribes the original narration → Claude performs fact-preserving rewriting (structure reorganization + language substitution; see feasibility analysis below).
3. Map recreation: Reverse-engineer event data (time / location / unit / trajectory) from map frames, then rebuild animations in Google Earth Studio or Mapbox — this is the heaviest-lift环节 and also your moat.
4. Assembly: Edit and compose (see the two routes below).

Only two manual intervention points throughout: topic selection approval and final QC. Everything else is fully automated.

Two Routes at the Assembly Layer: Draft Direct-Write vs. MCP-Driven

As of 2026, a new option has emerged at the assembly stage: agents can now control JianYing (CapCut) and Premiere via MCP to do editing, making “fully automated video commentary” technically achievable. But the engineering properties of the two routes differ significantly, and picking the wrong one will bottleneck you at scale:

The assembly needs of map-narrative videos fall entirely within the former’s capability: track layout, subtitle bars, avatar cards, transitions, and beat-matching are all templated components. So this pipeline’s assembly layer goes with draft direct-write — generate the draft JSON + TTS audio track + asset placeholders offline, open in JianYing to preview and tweak, with human QC happening at this step. MCP-driven editing is reserved for post-scale refinement needs (e.g., making an enhanced remake of a viral hit).

One heads-up: MCP editing pushes “fully automated” from 95% to 100%, but assembly accounts for only about 5% of total pipeline workload — map animation production (60%) and copy skeleton (20%) are the real leverage points. Tool hype easily misdirects effort toward the smallest slice.

Genre Selection: Four Candidates

Recommended combo: 70% disaster investigations + 30% expedition accidents during the cold-start phase. Rationale: disaster investigation reports sit at L0 and are public (NTSB/AAIB report PDFs are freely downloadable), so you’re effectively building L1 on top of L0 with the cleanest copyright position. Expedition accidents piggyback on the niche Mysterious Garden already validated, but with topic错位 (they do hiking and cave diving; you do aviation and mining disasters). Save classic battle walkthroughs for after the account gains traction — that’s the highest-traffic but most cutthroat segment.

The Core Question: Is Video Recreation + English-to-Chinese Rewriting Feasible?

I’ll split this in half.

Map recreation: feasible, and this is precisely what you should do. Map animations themselves aren’t copyright-protected (facts and geographic information aren’t protected; expression is). As long as you rebuild the visualization yourself — the same battlefield, your own icon language, your own animation pacing — it’s a legally independent work. Workload: a medium-scale battle’s map animation takes 40–60 hours manually in After Effects; with “frame extraction → data reverse-engineering + procedural generation,” you can compress that to under 10 hours. That’s exactly the value of an automated pipeline. And Mysterious Garden has already proven that Chinese audiences will pay (in views) for self-made maps.

English-to-Chinese copy translation + LLM rewriting: grey zone, doable but with a ceiling. The right of translation is part of copyright; line-by-line translation, even in a different language, still falls under infringement. LLM rewriting can create distance, but the legal standard is “whether the expression is substantially similar,” not just word-level overlap — if the narrative structure, entry points, and fact-selection order all mirror the original (frame-by-frame correspondence), no amount of rewriting changes the fact that it’s a “derivative work,” and unauthorized derivatives are infringement. In practice, no one in the Chinese space has been pursued (花花有话说’s 170K likes and thriving), and platforms largely ignore it. Risk assessment: near-zero risk at small scale; once you reach hundreds of thousands of followers, the probability of the original channel taking notice becomes real; after you activate ad monetization (MCN affiliation / sponsorships), commercial use significantly raises the probability of enforcement.

So my recommendation is a hybrid strategy: Don’t translate the narration line by line. Instead, do “fact-skeleton reconstruction” — extract a fact清单 from the original (timeline / locations / persons / event sequence; these aren’t copyright-protected), then write your own Chinese copy using your own narrative structure, with the LLM handling only fact-checking and polishing. This adds ~30% workload versus pure translation, but it shifts you from “repost with translation” to “independent creation,” and the ceiling is entirely different. Mysterious Garden’s approach, with costs compressed by automation, can undercuts theirs.

The Final Piece of the Commercial Loop

The logic of follower growth and ad revenue: map-narrative content has naturally high completion rates (suspense hooks + timeline progression), and Douyin/Bilibili’s recommendation algorithms favor high-completion long videos (5–15 minutes) with traffic boosts. This segment’s follower value per capita is higher than broad-entertainment content. Advertiser profiles: military/disaster → male-skewed categories (gear / 3C / gaming); expedition → outdoor brands. The ad-revenue threshold sits between 100K–500K followers. At Mysterious Garden’s growth rate (1M in 3 months), expect 6–12 months to reach that range.

One number still needs validation: I estimate that L1 channels produce 15–25 qualifying videos per week (P25+ view percentile, safe-zone genres). Existing Chinese players (Mysterious Garden + repost accounts) collectively consume roughly half that. The remaining supply is your pipeline’s raw-material feed — this estimate needs to be calibrated with real data after two weeks of Windmill monitoring. That’s also the first module on the entire pipeline that should be deployed first.