playbook
I Taught My AI to Run 9 Projects
6 Mar 2026
A storytelling playbook: narrative spine, identity framing, and content plan — researched across 8 master builders and the AI content landscape.
Projects 9 active
Reports 53 in 5 wks
Dunbar 71/150
Risk No narrative
STRONG
"I taught my AI to run 9 projects. Here's what happened." Narrative spine backed by 53 live reports, 2 paying clients, 3 pilots, persistent memory across 71 contacts. Identity frame: Founder. Builder. NOT consultant. Ship LinkedIn carousel first, then 7-day sprint.
→ The output is elite. The story hasn't been told. This is the playbook.
Eric San
learn
Learn: MediaPipe Hand Panels + WebGPU/TSL
3 Mar 2026
Reverse-engineering Berto's hand-tracked panel demo. Architecture is reproducible; final quality depends on live webcam calibration and smoothing constants.
Donna Partial
Eric Calibration pass
Time 60-120 min MVP
Blocker Live webcam tuning
BOTH MASTER
Pipeline is clear and buildable now. Donna can implement the WebGPU + TSL + MediaPipe architecture quickly; Eric only needs a short live-camera pass to tune jitter filtering and gesture feel for production quality.
→ Hard part is not shaders; it is stable gesture mapping and smoothing.
Analyst: Eric San
deep research
Rattan & Bamboo: Vietnam Trading Deep Dive
3 Mar 2026
Market sizing, competitive positioning, and destination analysis for Vietnam-based rattan/bamboo export trading — UAE CEPA advantage, Indonesia rivalry, trader margins.
Global TAM $7.5B
Vietnam 2024 $724M exports
UAE Tariff 0% (CEPA)
Death Indonesia 42% share
STRONG
Start with UAE basketwork/décor bundle. Rattan/bamboo offers credible first market: CEPA 0% tariff (Feb 2026), fragmented supply (893 craft villages), clear margin path (15–25% wholesale → 30–40% private label). Indonesia dominates raw material but Vietnam wins on compliance + CEPA access.
→ CEPA arbitrage + craft village aggregation = trader wedge. Not resource competition with Indonesia; compliance/access competition.
Vincent Li (Vietnam Trading)
deep research
Sourcing Basketwork from Vietnam
3 Mar 2026
16 vegetable materials scored, 5 destinations ranked, full COGS, certifications by market, buyer personas, and sales strategy for Vietnam basketwork export trading.
Materials 16 evaluated
Concentrate 5 selected
UAE Tariff 0% (CEPA)
Death Mold in transit
STRONG
Concentrate on water hyacinth + seagrass + rattan + bamboo + coir. These 5 share Vietnam supply dominance, established craft village infrastructure, and proven export demand. Water hyacinth is the standout: zero raw material cost, dense/nestable (no volumetric freight penalty), eco-upcycling premium. 80% of craft villages cannot self-certify — the compliance bridge IS the moat.
→ Water hyacinth = best fit material. Compliance bridge = structural moat. Mold = the only kill variable.
Vincent Li (Vietnam Trading)
tech
Agent Compute & Data Claims
3 Mar 2026
Proxy APIs, grey market Claude keys (Taobao), DGX self-hosting economics — stress-testing Penny's claims on cost optimization.
Taobao keys ~54% cheaper
Max plan 93% savings
DGX break-even 2-3 years
ToS risk Account bans
CONDITIONAL
Proxy APIs worth trialing for non-sensitive workloads. Taobao grey market is technically real (~$1.375/M tokens vs $3 official) but carries material ToS/account ban risk. DGX only viable at >5M tokens/month sustained. Official Max plan dominates for Eric's current usage.
→ Prompt caching (official, 90% savings) > Max plan (high volume) > Proxy APIs (B2B only) > Grey market (experimental only).
Analyst: Eric San
learn
Learn: Vercel agent-browser
3 Mar 2026
AI-first browser automation: snapshot + @ref for ~80–93% token savings. Donna reproduced open → snapshot -i → click/fill/screenshot → close with zero blockers.
Donna Mastered
Eric Cleared
Time < 15 min to test
Blocker None
verdict-strong
Donna can run agent-browser on command. Install via npx or global; open, snapshot -i, click/fill by ref, screenshot --annotate, close. Fits research and form-filling flows with low token usage.
→ Snapshot + refs replace full HTML in context; same primitives as Playwright but agent-optimized.
Analyst: Eric San
learn
Learn: ANE Training — Cracking Open Apple's Neural Engine
3 Mar 2026
Reverse-engineering Apple's locked-down Neural Engine for training. Donna reproduced on M4 Max: loss 0.500→0.001, 11.34 TFLOPS peak. Roemmele's ZHC@Home hype debunked.
Donna Reproduced
Eric M4 Max ready
Time < 2 hours
Blocker Private API risk
verdict-strong
ANE training works — ZHC@Home doesn't. maderix's reverse engineering is the real breakthrough (first-ever ANE training). Roemmele's distributed compute claim fails on 5 independent constraints. ANE efficiency (6.6 TFLOPS/W, 80× A100) strengthens Mac inference investment thesis.
→ For training: GPU/cloud. For inference: ANE is untapped. The 38 TOPS marketing = 19 TFLOPS real FP16.
Analyst: Eric San
learn
Learn: Real-Time LLM Fine-Tuning — Signal vs. AI Slop
3 Mar 2026
Roemmele's follow-up code is fabricated (model.to('ane') doesn't exist). The real path: MLX LoRA on GPU reprogrammed model identity in 10.5 seconds.
Donna Mastered
Speed 19.1 sps
Time 10.5s
Blocker None
verdict-strong
Fake code, real capability. Roemmele's ANE + PyTorch snippets are fiction. MLX LoRA on Apple GPU works today: 200 training steps in 10.5s, full identity shift from Qwen to Donna. Trivial to implement (~40 lines). Applications: per-client personality, domain fine-tuning during onboarding.
→ pip install mlx-lm + 40 lines of code. Fine-tuning beats RAG for personality/tone; RAG wins for knowledge retrieval.
Analyst: Eric San
learn
Learn: AI Short Drama Script Creation
01 Mar 2026
Reverse-engineering a highly structured, 7-step Claude Code workflow for creating monetizable, industry-standard short drama scripts.
Donna Mastered (Skill built)
Eric Taste assessment
Time < 30 min to test
Blocker Human emotional judgment
DONNA MASTERS
The artifact is essentially an AI agent prompt architecture. By extracting the core industry theory (Golden 5-seconds, Paywalls, 30-60-30 pacing) and encoding it into a Cursor Skill, Donna can now execute this exact 7-step process on command.
→ Micro-drama scripts are not literature; they are precision-engineered dopamine delivery systems. The value is in the pacing constraints, not the prose.
Analyst: Eric San
strategy & community
The Architecture of Elite Rooms
26 Feb 2026
Historical and modern anatomy of invite-only elite circles: what they got right, failure modes, indirect monetisation, and a 90-day execution roadmap for The Table.
Junto Club 38 years
Tiger 21 fee $30K/yr (facilitation, not membership)
YPO Forum 8–10 max, 100% nominated
Death Scale + overexplaining
STRONG — BUILD IT NOW
The concept is historically correct and the instincts are right. Every room that survived did so by resisting scale. Every room that died chased it. The economic thesis is sound: host authority compounds over 2–3 years into deal flow, advisory roles, and reputational leverage that directly accelerates any capital-intensive venture.
→ The best rooms are invisible. The Table's refusal to explain itself is not mystique — it is the mechanism that prevents quality decay.
Analyst: Eric San
tech research
Cursor Sub-Agent Setup
28 Feb 2026
How to set up and use Cursor subagents: Task tool, built-in Explore/Bash/Browser, custom .cursor/agents/, and when they make sense for your workflow.
Maturity Production (2.4+)
Cost Per-subagent context
Risk Model routing / no recursion
ADOPT
Already in use; no migration. Built-in Explore/Bash/Browser run automatically. Add optional custom agents (verifier, debugger) under .cursor/agents/ with YAML + prompt. Fits /deep* research, project checkout, shipios. Sourcy: yes for Cursor-side work, not for WA bot runtime.
→ Subagents are the right tool for context-heavy and parallel work. Custom setup is optional; 1–2 hours for verifier + one specialist.
Analyst: Eric San
legal & negotiation
Cross-Border Legal Exposure: Italy to Hong Kong
26 Feb 2026
R2: Critically assessed. Cap 319 civil enforcement confirmed, but Red Notice overstated and settlement can be negotiated lower.
Target €45–55k
Extradition None (HK ↔ Italy)
Civil Risk Cap 319 (with defenses)
Death Ignoring judgment + HK assets
NEGOTIATE & SETTLE
Settle — but negotiate harder. Cap 319 civil enforcement is real (confirmed). But Red Notice threat is overstated for private disputes (INTERPOL Art 83). Enforcement costs €20–40K — giving the friend leverage to target €45–55K (vs. R1's blanket €80K). 12 sources, 6 corrections from Gemini R1.
→ Cap 319 enforcement is real but not automatic. INTERPOL prohibits Red Notices for private disputes. Cross-border enforcement costs create negotiation leverage.
For: Penny Yip · Analyst: Eric San
market & strategy
Agent-to-Physical Orchestration
25 Feb 2026
The actuation graveyard (FB M, Magic, GoButler) and why FieldCheck's camera-verified fulfillment is the missing moat for agentic APIs.
Agent TAM $10.75B
FSM TAM $5.1B
Graveyard 5 dead
Death Ops cost of fulfillment exceptions
STRONG MOAT — VERIFICATION IS THE MISSING PIECE
High value, high fatality risk. Atoms don't return 200 OK. If an agent triggers a physical task and the supplier ignores the ping, human operators must fix it, destroying margins (The Death Cost). Avoid the B2C 'universal API' illusion that killed Magic. FieldCheck solves the exact problem that killed horizontal API startups: fulfillment verification. By putting a camera in the loop, FieldCheck replaces human QA ops with VLM analysis, closing the loop so agents know work was done correctly.
→ FieldCheck solves the exact problem that killed Facebook M and Magic: fulfillment verification.
Analyst: Eric San
deep thesis
The Intelligence Displacement Thesis — Critical Assessment (R2)
23 Feb 2026
Falsifiable framework, 4 historical displacement waves, primary research (JOLTS, MIT, SaaStr, India IT), and micro-level stories. R2 adds 24 new sources and reformed verdict.
WC openings P&BS -21.8%, Finance -25.1%
AI pilot fail 95% zero ROI (MIT)
SaaS credit $600-750B exposed
Sources 34
Risk Financial contagion via SaaS credit cascade
UNCERTAIN — WRONG ON WHEN, RIGHT ON HOW
Directionally correct on the what, wrong on the when and how. Primary research confirms white-collar displacement (JOLTS: P&BS -21.8%), entry-level collapse (dev hiring -67%), and SaaS credit stress ($17.7B distressed, most since Oct 2008). But 95% of enterprise AI pilots fail, GPT-5 showed 1-point benchmark gain (S-curve), and M2V is recovering. The thesis may be wrong about labor timing but accidentally right about financial contagion — SaaS private credit doesn't need mass unemployment to blow up.
→ SaaS private credit ($600–750B) doesn’t need mass unemployment to blow up — the financial cascade precedes the labor one.
Analyst: Eric San
thesis
Decision Intelligence at Infinite Leverage
15 Feb 2026
As agents climb from intern to cofounder, the highest-leverage output is decision-making. The four inputs — IQ, taste, context, data — become the defining assets of the AI economy.
Consulting TAM US$500B+ disrupted
IQ Cost $36→$0.40/Mtok in 18mo
Palantir 56% growth, $70B val
Taste Last human scarcity
Risk AI develops taste (5–10yr)
THESIS: DECISIONS ARE THE PRODUCT
Decisions are the highest-leverage output humans produce, and AI is repricing every input. As intelligence commoditizes (10x/year cost decline), the remaining decision components — taste, private context, expanded data — capture all marginal value. The $500B+ consulting/advisory market is being disrupted. Eric is positioned: Donna is the context container, the research engine is the data expander, claw.degree measures taste, @ericsanio is the proof-of-work.
→ As intelligence commoditizes, taste and private context capture all marginal value in decision-making.
Analyst: Eric San
thesis
How to Build Applications with Agentic Backends
14 Feb 2026
The winning architecture is boring — 50+ players, 7 live debates, 40+ production cases. What actually ships.
Pattern While-loop + tool_use
Players 50+ mapped
Production 40+ cases
Cost $0.05–0.30/turn
Risk Latency — 2–15s per tool cycle
THESIS: LOOP > FRAMEWORK
The best agentic backend is 200 lines calling Claude with your tools. Every product above $50M ARR uses this pattern. Frameworks add value only for enterprise orchestration (durability, audit trails, multi-agent). 90% of builders don’t need them. The moat is domain-specific conversation design, not the agent architecture. Cursor CLI for personal automation; Claude API while-loop for multi-user products.
→ Every product above $50M ARR uses the same pattern: a while-loop calling an LLM with tools. Frameworks are optional.
Analyst: Eric San
deep thesis
The Token Orchestration Thesis
21 Feb 2026
As token budgets surpass salaries, the human who deploys AI captures asymmetric value — but only under specific structural conditions.
Leverage 1 human → 100 drones
History 60-90% gains to capital
Requirement Ownership
CONDITIONAL (OWNERSHIP REQUIRED)
The orchestrator only captures asymmetric value if they OWN the outcome. 250 years of history shows automation gains are captured by shareholders, not workers. Employees who get 10x more productive have their surplus extracted. You must be a founder, freelancer, or equity holder to benefit.
→ 250 years of automation history: 60–90% of productivity gains are captured by capital, not labor. Ownership is the prerequisite.
Analyst: Eric San
deep investment
Agentic Theses Investment Research (R4)
22 Feb 2026
The agentic shift is real but consensus-long. Edge exists in infrastructure timing, valuation gaps, and sensor-driven physical asset arbitrage.
Assets 100 mapped
Portfolio $50K allocated
Edge Structural + Timing
Sensors 3 live quant models
EXECUTE PORTFOLIO
R4 massive expansion. 24GB+ compute history mapped. 4 categories: Hardware SKUs, Equities/Cooling, Web3/DePIN, Private SPVs. Programmatic sensors built. The agentic shift is consensus-long — edge is in WHERE value accrues (landlords + power), not THAT it accrues. $50K allocated across Sovereign Hardware, DePIN, Pre-Seed, Cooling Equities.
→ The agentic shift is consensus-long — edge is in WHERE value accrues (infrastructure, power), not THAT it accrues.
Analyst: Eric San
deep investment
Roblox (RBLX) — Position & Investment Analysis
27 Feb 2026
Competitive position, consensus map, unit economics. Leading UGC platform; no path to profit yet, edge unclear.
DAU 144M (Q4 25)
Bookings FY25 US$6.8B
P/S ~9×
Death No profit path; metric/regulatory risk
CONDITIONAL — MONITOR
Leading UGC platform; no clear edge. Roblox leads by DAU and engagement vs Epic/Fortnite. High institutional ownership, Buy consensus, but unprofitable and faces short thesis (Hindenburg, Potato Capital). Edge unproven — learn and monitor; small conviction only unless you have a specific view on metric quality or path to profit.
→ Position strong; investment edge unclear. Monitor operating leverage and ad monetization in 2026.
Analyst: Eric San
thesis
The Personal Agent Is the New Email Address
11 Feb 2026 (R3)
Refined thesis: personal agents (OpenClaw/Donna model) as identity technology, not the generic “AI agent” noise. Identity techs reach universality. Tools plateau.
Framework Identity vs Tool
Runners 10–50K globally
Era Pre-Hotmail (~1993)
Hotmail Moment ~2027–2029
50%+ Adoption ~2033–2037
Risk Trust delegation is novel
THESIS: IDENTITY → UNIVERSALITY
Identity technologies reach universality. Tool technologies plateau. Personal agents (persistent, contextual, representing you) are identity — not tools. 10–50K runners today = ARPANET-era email. The generic “AI agent” market is 90% agentwashing (130 genuine out of thousands). Ignore that noise. The personal agent adoption curve follows email/smartphone, not VR/IoT. Hotmail moment ~2027–2029. Build for the 10–50K early adopters now.
→ Identity technologies reach universality; tool technologies plateau. Personal agents are identity, not tools.
Analyst: Eric San
deep investment
Secondhand Compute as Inference Infra
22 Feb 2026
The arbitrage between consumer electronics pricing and AI infrastructure value. RTX 3090, M2 Ultra 192GB, A40 — mispriced nodes, no-lose prices.
RTX 3090 Conviction (<$800)
M2 Ultra High Conviction (<$2,000)
RTX 4090 No Edge
Window 18-24 months
STRONG BUY (SPECIFIC SKUS)
Edge: structural + horizon. RTX 3090 delivers 93% of 4090 bandwidth at 40-55% cost. M2 Ultra 192GB used = $1,800-2,200 vs $6,000+ new. Window is 18-24 months before next-gen creates illiquidity. Sub-24GB consumer cards have zero edge.
→ RTX 3090 delivers 93% of 4090 bandwidth at 40–55% cost. Consumer electronics pricing lags AI infrastructure value.
Analyst: Eric San
deep investment
Private Inference as a Service
22 Feb 2026
The $0.89/M to $12/M margin hiding in regulated industries. Healthcare, legal, and finance cannot use shared cloud AI.
Premium 3-8x VNet, 10-90x Air-gap
Floor $0.14/M DeepSeek myth
Real Cost $0.45-1.78/M blended
Cloud Equity Fully priced
OPERATIONAL PLAY, NOT EQUITY
Edge: information + structural. The $0.14/M DeepSeek commodity floor is a myth for regulated sectors — they pay 10-90x premium because regulators REQUIRE it, not by preference. This is an OPERATIONAL play (build the service), NOT a financial investment, as cloud equity (AMZN/GOOG) is already consensus-long.
→ The $0.14/M DeepSeek floor is a myth — regulated sectors pay 10–90x because regulators require it, not by preference.
Analyst: Eric San
market & strategy
FairContract.io — AI Contract Negotiation for Early Partners
26 Feb 2026
Neutral AI that shows both parties their bias, maps divergence points, and proposes a middleground version. Novel framing — no one does exactly this.
TAM US$2.8B (contract analysis)
Margin 99.7% (US$0.09 COGS)
Gap No neutral tool exists
Death Trust barrier + ChatGPT free
CONDITIONAL — WEEKEND EXPERIMENT
The wedge is genuinely novel. No existing tool shows both parties their biased views alongside a neutral middleground. Pain is quantified (55B hours/year wasted, 73% co-founder disputes). Economics are beautiful (99.7% gross margin). But trust barrier is real — will people trust AI on legal matters? Ship in a weekend, test for 30 days, data decides.
→ No one builds the neutral third party. Every contract AI is adversarial. The cooperative framing is the wedge.
Founder: Eric San
deep market
Padel Operator in Hong Kong
25 Feb 2026
Venue setup, licensing, day-to-day ops — and whether the economics survive HK rent. Sweden’s 90 bankruptcies, 3 HK venues mapped, full CAPEX/OPEX model. 30 sources.
HK Courts ~13 (3 venues)
Court Rate HK$700–1,000/hr
CAPEX HK$1–1.75M (4 courts)
Death Rent >40% of revenue
Sources 30
CONDITIONAL — RENT IS THE ENTIRE BET
Only viable if you control the rent. HK has ~13 courts for 7.4M people — extreme under-penetration. UBS charges HK$1,000/hr and fills pop-up courts. But Sweden’s We Are Padel lost US$70M because rent > revenue, and HK rent is the world’s highest. The 6m ceiling requirement kills 80% of industrial stock. Find NT warehouse at HK$10–14/sqft or don’t start. Total commitment: HK$3–5M.
→ Sweden’s 90 padel bankruptcies prove one thing: if rent exceeds 40% of revenue, the venue dies regardless of demand.
Analyst: Eric San
R2 + critical
Cursor Skills Marketplace (R2 + Critical Assessment)
24 Feb 2026 (R2)
Full redo with 35 sources, startup graveyard, COGS table, and integrated /criticallyassess. V1 had 12/17 quality items missing. R2 stress-tests every assumption.
Discovery 958K visits/mo (SkillsMP, 18x growth)
Free supply 200K+ skills (all free)
Paid proof Apify $500K/mo (tools, not taste)
Death Free norm + AI closes taste gap
UNCERTAIN — WRONG PRODUCT SHAPE
Right demand signal, wrong product shape. Demand for skill discovery is explosive (SkillsMP 18x growth). But every platform at scale is free. Only marketplace paying creators (Apify $500K/mo) sells functional tools, not taste files. Sell outcomes (consulting, configured agents) or functional MCP tools — not static skill files. V1’s AEO/GEO framing rested on a vendor survey (Conductor, 97%) that Gartner contradicts (23% optimized).
→ Every skill marketplace at scale is free. Only platforms selling functional tools (not taste files) monetize.
Analyst: Eric San
Part 3: Non-Exclusive Category Discovery (R2)
17 Feb 2026 (R2)
Profile-first discovery: what do lawyer-connector-trust people succeed at globally? Then checked Vietnam demand. Found 3 new STRONG-fit categories beyond Part 2.
New Strong 3 categories found
VN Franchise $22B market
VN HNWI 19,400 (+98% in 10yr)
VN PE $2.3B / 141 deals
Death All 4 are same biz — execute as 1
4 STRONG FITS — CONVERGE TO 1 FIRM
M&A advisory, franchise brokerage, HNWI wealth advisory, and deal sourcing are not 4 businesses — they’re 4 revenue streams of one cross-border intermediary firm. Critically assessed R1 (which found “zero hidden gems”) and broke it: methodology was backwards (economy→founder instead of founder→economy). Profile-first method found franchise brokerage ($22B VN market), HNWI advisory (19,400 millionaires, fastest growth globally), and PE deal sourcing ($2.3B, 141 deals). All leverage who Vincent IS: trusted cross-border intermediary.
→ Start opportunity discovery from the founder’s profile type, not the target economy. Economy-first misses what the founder is built for.
Founder: Vincent Li
Donna-for-F&B — HK Restaurant WhatsApp CRM (R4)
18 Feb 2026 (R4)
Part I: Access-first GTM (7 opportunities, staffing #1). Part II: Deep market validation — TAM $2.8B, 8 competitors dissected, unit economics, GTM via own network.
Access OWNED (Black Sheep, Pirata, Arca) + PARTNERED (Bob)
Wedge WhatsApp-native agentic (nobody in mid-tier APAC)
Margin 58–89% gross at HK$1,500/mo
Risk Loyalty graveyard (Mogl/$35M dead) — must lead operational
Proof Loman $199–399/mo, 500+ restaurants paying for AI agents
CONDITIONAL — VALIDATE WITH 1 CONCIERGE PILOT
R4 adds deep market research. Agentic + WhatsApp-native + mid-tier HK = unoccupied position. Vynta (luxury) and Nuxa (orders) are closest but don’t match. Unit economics work (58–89% margin). But: restaurant loyalty is a graveyard, software is <1% of spend. Lead with OPERATIONAL automation (“I handle your WhatsApp”), not loyalty. Validate: 1 concierge pilot at Black Sheep or Pirata. Kill if 0/2 willing to pay.
→ Restaurant loyalty is a graveyard (Mogl/$35M dead). Lead with operational automation, not loyalty.
Eric San (product + access) · Bob (amplifier channel)
Blue-Collar AI
11 Feb 2026 (R3)
B2B vertical SaaS with hardware — construction safety, warehouse QC, field inspections, blue-collar ops
Dog-food Smilie packing QC
TAM US$6.2B→$18B
Construction 300K+ SG workers
GTM Warm network + 2 construction contacts
Margin 60–75%
Death AI inference cost + Protex AI ($36M, YC)
CONDITIONALLY STRONG
Do it — but sequence correctly. Construction emerges as largest potential scale vertical (300K+ workers, WSH compliance, BETC 70% grants). Protex AI ($36M) is a new competitor but uses fixed CCTV, not worker-level guidance. SafetyCulture ($2.7B) proves PLG safety inspection market. Sequencing unchanged: packing QC dog-food → repair.sg pilot → validate construction + cleaning + hotels in parallel.
→ Protex AI ($36M, YC) uses fixed CCTV; worker-level mobile guidance is the unoccupied position in blue-collar AI.
Founders: Wenhao Dong & Eric San
Part 2: Vietnam / HK Opportunity Screen
17 Feb 2026
10 categories screened across 6 research layers (VN forums, Threads, web, Google Trends, Reddit, LinkedIn). Founder-fit assessment for a cross-border operator.
Categories 10 screened
Survivors 3 advance
Sources 80+ queries, 6 layers
VN M&A 367 deals, $8.7B
Death Ops experience gap
SCREENING — 3 ADVANCE, 7 KILL
The three survivors leverage who you are, not what you could learn to do. Cross-Border Legal Advisory (#1, day-one executable), Authenticated Health Imports (#2, needs ops partner), Legal Tech SaaS (#3, needs tech co-founder). Killed: D2C imports, premium beverages, fashion, smart home, eco, concierge, standalone baby products. Pattern: anything requiring FMCG distribution, technical skills, or operational grind where the founder's profile is resume decoration, not competitive moat.
→ Every surviving opportunity leverages who the founder IS; every killed one requires what they lack.
Founder: Vincent Li
deep market
Vietnam Trading & Sourcing Business
22 Feb 2026 (R2)
A rigorous assessment of the Vietnam export trading business for a zero-experience founder. TAM, unit economics, graveyard, competitive landscape, GTM.
Macro Hostile (Tariffs)
Defect 15% rate wipes GM
Moat HK Network
CONDITIONAL (JOB FIRST)
Profile mismatch + hostile macro environment. 46% tariff risk and 2.5x sourcing costs. A silent partner removes the #1 kill switch. HK network is a real moat but only activates with a specific product story. Take a job at a sourcing firm to learn logistics and build network on someone else's dime first.
→ A silent partner removes the #1 kill switch for first-time physical goods operators: operational inexperience.
Founder: Vincent Li · Analyst: Eric San
Video-to-Text as a Service
17 Feb 2026
Batch-process TikTok/IG Reels into structured text at scale, sell as API to AI agents. Gemini batch kills the margin before you start.
Gemini 2.5 Batch $1.67/1K videos
Twelve Labs $107M raised + MCP
Addressable $2–8M (thin)
Dog-food PCRM needs it
Death Google batch = same price for everyone
SKIP API — BUILD THE SKILL
Don’t build the API business. Build the PCRM skill. Gemini 2.5 Flash batch at $1.67/1K videos — same price for everyone. Twelve Labs ($107M) already has MCP server for agents. Apify already has Video-to-Text MCP. The real value is in the judgment. Build tiktok-research + ig-reels-research as Donna skills. Cost: $0.20 per 100-video query.
→ Economies of scale don’t work when the commodity provider gives everyone the same bulk pricing.
Founder: Eric San
tech research
Cursor → Claude Code Migration
22 Feb 2026
Keep Cursor as primary IDE, add Claude Code as power tool for heavy autonomous work. Target: cut AI spend from ~$3,400/mo to ~$1,400/mo.
Spend $3.4K → $1.4K/mo
Portability 60% of PCRM
Gap Visual/CRM workflows
HYBRID WORKFLOW
Full switch NOT recommended — terminal-only is a downgrade for visual/CRM work. Augment with Claude Code Max ($200/mo) for heavy report sprints to save API costs. 40% of PCRM (24 commands, 21 skills) requires effort to port.
→ Terminal-only AI coding is a downgrade for visual work. Hybrid beats full migration: IDE for daily ops, CLI agent for heavy sprints.
Analyst: Eric San
thesis
The Agent Tooling Landscape
14 Feb 2026
Cursor Agent vs Claude Code vs Claude Agent SDK vs OpenAI Agents vs OpenClaw — taxonomy, market scoreboard, and the “tool < agency” roast
Cursor $1B ARR, $29.3B val
OpenClaw 188K stars
OpenCode 2.5M devs
5 Categories IDE / CLI / SDK / Framework / Runtime
Death Interface layer gets absorbed (Windsurf, Supermaven)
LANDSCAPE ANALYSIS
Five categories, not one market. IDE agents (Cursor), CLI agents (Claude Code), SDKs (Claude/OpenAI), frameworks (LangGraph), and runtimes (OpenClaw) serve different jobs. Interface-layer tools die in consolidation. Model + identity layers are durable. “Tool < agency” is 60% cope, 40% insight — agency is expressed through tool affordances, not independent of them. MCP maturity is the unlock.
→ Five categories, not one market. Interface-layer tools die in consolidation; model + identity layers are durable.
Analyst: Eric San
product research
AI Tools for HK Small-Mid Law Firms
18 Feb 2026 (R2)
What to build for the 88% of HK law firms that big AI vendors have ignored. Confidentiality is the moat. 800 firms, 1,800 barristers.
Unserved 88% of firms
Market 800 firms, 1.8K barristers
Moat Confidentiality
STRONG NICHE
Big AI vendors ignore local private stacks. 88% of the market needs locally-hosted or privacy-guaranteed AI tools. Build specifically for HK legal workflows (bilingual OCR, case bundle summaries) with absolute data sovereignty.
→ 88% of HK law firms are unserved by big AI vendors. Confidentiality is the moat for local private stacks.
Analyst: Eric San
claw.degree — AI Agent Grading
11 Feb 2026
HubSpot Website Grader playbook for AI agents — paste agent URL, get a score
Gap No consumer-grade scoring exists
Margin 92%+
COGS $0.05/eval
Domain $8/yr available
Death Score credibility
STRONG — BUILD NOW
Fills the widest gap in the agent infrastructure stack. Academic benchmarks serve researchers. DevOps tools serve engineers. Nobody serves end-users asking “is this agent good?” HubSpot Website Grader drove 10M+ leads. 92%+ margin, $0.05/eval COGS. Natural eval layer for Agent Elo. Start with objective dimensions only.
→ Nobody serves end-users asking “is this agent good?” The consumer-grade agent scoring gap is wide open.
Founder: Eric San
gtm access
Catherine's Counselling Network — GTM Access
22 Feb 2026
Catherine's master's cohort = warm double-sided access: supply (counsellors) and demand (clients via word-of-mouth) simultaneously. No incumbent serves HK new graduates. Directory + peer consultation as first wedge.
Supply access PARTNERED (cohort)
Demand access REACHABLE (word-of-mouth)
Top opportunity HK counsellor directory
Fast wedge Peer supervision group (week 2)
Risk PARTNERED base rate 70% fail 2yr
BUILD: DIRECTORY + SUPERVISION LOG
Double-sided warm access is rare. Counsellors need clients; clients can't find Cantonese counsellors. The HKPS supervision log tracker is the killer feature — every new graduate needs it, no global tool serves it. Fastest first revenue: Catherine runs a paid peer consultation group in week 2 (zero code). Directory v0: 5 profiles + Calendly in 1 week.
→ Double-sided warm access (supply + demand simultaneously) is the rarest and most valuable GTM position.
Analyst: Eric San · via Catherine
product research
HK Counselling Platform — What to Build
22 Feb 2026
Feature-by-feature product research for two user types: newly-qualified HK counsellors (Catherine's cohort) and their clients (word-of-mouth self-referred + public). P0/P1/P2/P3 prioritisation with competitive mapping.
P0 features 3 (counsellor side)
P0 features 2 (client side)
Killer feature HKPS supervision log
AI feature Session note assist (P2)
Risk PDPO compliance required pre-launch
BUILD ORDER CLEAR
Supervision log tracker is the only feature no global tool provides for HK counsellors. Pair it with directory + booking as the v1. Every other feature — AI notes, journaling, self-assessment — is phase 2. Client side needs: Cantonese search, fee transparency, video intro, booking. Don't build AI triage without a crisis human fallback protocol.
→ Build the wedge no global tool provides. For HK counsellors, that’s the HKPS supervision log tracker.
Analyst: Eric San · via Catherine
Creator Cohort & Media Agency — HK Creator Economy
12 Feb 2026
AI automating 80% of agency overhead. Community is the moat, not intermediation. First APAC creator cohort.
Community Gap 230 replies, no hub
HK Influencer Ad ~US$70M/yr
APAC First No cohort exists
Agency AI automating function
Death 3% commission disruptor
CONDITIONAL — COHORT NOT AGENCY
The cohort is the real idea. The agency model is structurally threatened. HK creators hungry for community (230 replies on Threads, no hub). Zero APAC cohort programmes found across 6 platforms. But AI is automating 80% of agency overhead (Reacher YC S25, bountiesdotwork, 3% commission disruptor). Don’t build a traditional agency — build a creator community with embedded brand access. Monetise through paid tiers, not commission.
→ AI is automating 80% of agency overhead. Don’t build a traditional agency — build community with embedded brand access.
Founder: David Li
The Knowledge Call Network
15 Feb 2026
Networking platform where 1-on-1 recorded calls produce transcripts that become searchable, gated knowledge assets — the supply-side unlock for decision data.
Dog-food 21 PCRM transcripts
Demand Admond TRV “deeper 1-1s”
Analog Tegus library → $930M acq
COGS $0.01/min transcription
Death <20% transcript access rate
CONDITIONAL — FEATURE LAYER
Build as feature layer, not platform. The call is the value, the transcript is the data asset. Networking platforms have a brutal graveyard (Clubhouse $4B→dead, Lunchclub fading). Don’t build a new destination — build a transcript gating/search layer for existing communities. Start with PCRM’s 21 transcripts, expand through TRV’s members. Test: do 20%+ of members actually search/access others’ transcripts?
→ Knowledge marketplace platforms have near-100% failure rate as standalone destinations. Build as a feature layer.
Founder: Eric San
The Decision Data Marketplace
15 Feb 2026
As agents climb from intern to cofounder, the bottleneck shifts from intelligence to private context. “Uber for life knowledge” — and the $3.8B market that already exists.
Expert Networks US$3.8B, 16% CAGR
Decision Intel US$18B (2025)
Dog-food PCRM research engine
Death Supply-side economics — individual knowledge near-zero value
CONDITIONAL — PROTOCOL, NOT PLATFORM
The thesis is correct. The timing is wrong. The product shape is wrong. Experiential data is the one data layer AI can’t crawl. But every personal data/knowledge marketplace has failed — unit economics break at the individual level. Build as a protocol feature for personal agents (Shape C), not a standalone marketplace. Revisit when agents hit ~1M users (~2027–2029).
→ Every personal data marketplace has failed. Unit economics break at the individual level. Protocol beats platform.
Founder: Eric San
tech research
Claude Code Agent Teams
22 Feb 2026
Multi-agent orchestration in Claude Code: architecture, feasibility, and mastery guide. Stress-tested through critical assessment.
Pattern Peer-to-peer messaging
Cost 2x Token overhead
Bugs VS Code + tmux delivery
TRIAL
Sound architecture, buggy execution. Best for bounded experiments on greenfield modules with clear file ownership. Keep Cursor Task subagents as the daily workhorse until messaging delivery bugs are resolved.
→ Multi-agent architecture works in theory; message delivery bugs break it in practice. Bounded greenfield experiments only.
Analyst: Eric San
tech
Math Diagram Generation for HKDSE 3D Geometry
24 Feb 2026
Spike-tested SVG and TikZ pipelines for generating exam-accurate 3D geometry figures from structured diagramSpec data. First-party evidence from 3 DSE14 test cases.
SVG Zero deps, good quality
TikZ Academic-grade, CM fonts
AI images 73.9% accuracy (avoid)
Sources 10
TRIAL — SVG FIRST, TIKZ FOR POLISH
Both SVG and TikZ pipelines work for structured diagramSpec. Direct SVG generation is the pragmatic first step: zero dependencies, full layout control, and the LLM generates correct geometry from well-structured specs. TikZ produces more academic output (Computer Modern fonts, proper math symbols) but adds a compilation step. Avoid AI image generation for geometric accuracy. Recommended: SVG for speed, TikZ for teacher-facing polish.
→ Algorithmic SVG beats AI image generation for geometric accuracy (73.9% AI benchmark). Structured specs are the key input.
Analyst: Eric San
tech
PhotoDrop Packaging — Can We Ship Zero-Dep?
25 Feb 2026
Feasibility of packaging a Node.js photo gallery (sharp + exiftool + cloudflared) into a zero-dependency distributable for HK/TW photographers. 18 sources.
Single Binary FAILS (sharp)
npm/npx WORKS (2-4 hrs)
Electron WORKS (2-3 days)
Risk Over-engineering before validation
TRIAL — NPM FIRST
True zero-dep requires Electron (2–3 days + $99/yr Apple Dev ID). Overkill for a free Threads giveaway. Publish to npm as npx photodrop — sharp + exiftool auto-download, cloudflared ephemeral tunnel needs zero account. Photographer runs brew install node cloudflared once, then npx photodrop /path each time. Validate via Threads; upgrade to Electron only if traction proves demand.
→ Sharp’s native .node binaries kill every single-binary compiler (SEA, Bun, Deno, pkg). npm is the practical path.
Analyst: Eric San
E-E-A-T SEO Agency — Conversion-Driven Digital Marketing
15 Feb 2026
WordPress CMS, niche content, technical audits, industrial backlinks & digital PR — focused on driving conversions, not just rankings
SEO TAM US$75B, 17% CAGR
GEO/AEO US$886M, 182% YoY
Solo Margin 65–80%
AI CTR Drop −61% on AIO queries
Death Content labor + AI pricing compression
CONDITIONAL — PIVOT TO GEO/AEO
Right thesis, wrong vehicle. E-E-A-T focus is correct — Google rewards genuine expertise. But traditional SEO retainer is the shrinking part: AI tools ($500–2K/mo) compress pricing, AI Overviews eat organic traffic (−61% CTR), 56% zero-click searches. The better play: pivot to GEO/AEO “AI Search Visibility” — pre-category (US$886M, 182% YoY), LLM traffic converts 2x, no dominant player. Dog-food on research.ericsan.io, collaborate with Ronald Yao.
→ Traditional SEO is shrinking (−61% CTR on AIO queries). GEO/AEO is pre-category at 182% YoY.
Founder: Eric San
OpenClaw-as-a-Service
11 Feb 2026
PythonAnywhere-style managed hosting for OpenClaw agents — chat-triggered deployment
Dog-food Eric runs Donna on OpenClaw
TAM US$50–200M (OpenClaw-specific)
Validation SetupClaw $20k/wk
Stars 182K+ GitHub
Margin 40–60% Y1
Death OpenClaw launches official hosting
CONDITIONAL — MANUAL FIRST
Start with manual concierge, not a platform. SetupClaw ($20k/wk) and 182K GitHub stars prove demand. Chat-triggered UX is genuinely novel. But ClawCloud Run (official ecosystem platform) could add 1-click OpenClaw deploy anytime. Test with 10 manual setups at $50–150 before building. Build the platform only after 50 manual setups.
→ Start with manual concierge before building a platform. Official ecosystem hosting could absorb you anytime.
Exploring: Conrad Ho & Eric San
ola.tech — Can Used Phones Become Agent Infrastructure?
16 Feb 2026 (Consolidated)
Four rounds of research + critical assessment. The category is real (OpenClaw 196K stars, Doubao sold out), but the specific product has zero validated demand. $400 experiment, not $5K bet.
OpenClaw 196K stars, 600K downloads
Search “AI agent phone” = near-zero
Margin 29–65% (tier dependent)
Gap No one ships agent phones from SZ
Death Pre-config moat evaporating weekly
CONDITIONAL — $400 EXPERIMENT FIRST
Promising category, unvalidated product. Consolidated: Agent-on-phone ecosystem exploding (PhoneClaw 193 stars in 5 days, BotDrop 124 stars, Droidrun €2.1M). But no buyer searching for “AI agent phone”, DIY is trivial, pre-config value collapsing. Doubao ($495) validates premium; no evidence budget end ($49–99) works. List 5 phones, post to communities, talk to 20 users. Kill if 0–1 interest; scale if 10+ sales.
→ Agent-on-phone ecosystem is exploding, but DIY is trivial and pre-configuration value collapses weekly.
Founders: Raymond Sze & Eric San
AI Domain Trading & Hyperfinancialization
11 Feb 2026
Domains as tradable assets in the age of AI agents, content proliferation, and liquidity chasing everything
Aftermarket US$800M–1.2B
.ai Sales $22M (27 six-figure)
ai.com $70M record sale
Flipping EV $800/yr (100 domains)
Death Capital lockup + illiquidity
SKIP TRADING — TOOLS MAYBE
Macro thesis is real, but “trade domains” doesn’t follow. Most domainers lose money. Platforms consolidated (GoDaddy $437M). Tokenized domains unproven (ENS: 22 txns/day). Agent identity infra being built by trillion-dollar companies. Only viable play: productize existing CLI as SaaS for AI builders. Side experiment at best.
→ Most domainers lose money. The only viable play is productizing tools, not trading inventory.
Founder: Eric San
scout
PicPet HK Launch — Threads Viral Strategy
17 Feb 2026
Social media research brief: ready-to-use HK Cantonese Threads posts for PicPet’s Hong Kong launch, backed by TW viral post analysis
TW DAU 400K
HK Posts 1 (ground floor)
Top TW Post 11.7K likes
Downloads 623K+
Rating 4.8 iOS
Risk iOS only, no Android
WINDOW OPEN — FIRST MOVER
HK launched Feb 11, zero localized content. Two ready-to-use Cantonese Threads posts (couples + friends angles) calibrated from the top-performing TW viral posts. CNY timing window. Pet invite codes are the viral loop.
→ HK Threads viral posts are 3–8 lines max. Setup + turn + close. Punchline in the last 2 lines.
For: Jovy Yu
Let your agent be usable by other agents and humans. Let the best agent win.
TAM US$47B by 2030
Reach ~US$0 today
GTM Dog-food first
Timing Pre-category
Death LLM judge cost at scale
CONDITIONALLY YES
Flag-planting side project, not primary focus. MCP standardisation + agent proliferation = clear demand for quality/routing layer. Timing is early — not enough agents to rank yet. Eric is building agents (Donna, avet, OpenClaw) = natural supply side.
→ Not enough agents to rank yet. Pre-category timing means flag-planting, not building.
Founder: Eric San
Agent Utilities — Phone & Primitives
11 Feb 2026
“Twilio for Agents” — provision phone numbers, email, identity primitives for AI agents to operate in the real world
Gap No “AgentPhone” exists
Analog AgentMail (YC S25)
CPaaS US$14.3B market
Ceiling $50–200K ARR solo
Death Twilio/AWS absorbs as feature
CONDITIONAL — MCP TOOL ONLY
Real gap, wrong shape for standalone business today. Agent infra stack has funded players for compute (E2B $32M), browser (Browserbase $300M), email (AgentMail YC), payments (x402) — but nobody owns phone. Gap is real. But phone is centralized (carriers + Meta), Twilio can absorb in one sprint, and market is 10–50K runners. Ship open-source MCP server (10 hrs), bundle with OpenClaw hosting. Don’t build a company.
→ The AgentPhone gap is real, but Twilio can absorb it in one sprint. Ship open-source MCP, don’t build a company.
Founder: Eric San
avet — Agentic Vetting
8 Feb 2026 · v2
AI-powered qualitative member screening for community gating — platform-agnostic
TAM US$16B (IDV) + US$14B (moderation)
Reach US$50–200M
GTM Discord bot → Top.gg
Margin 95%+ (Discord/TG)
Ceiling US$36–120K ARR solo
Death WhatsApp API 30–45x COGS
WEEKEND EXPERIMENT
Upgraded from weekend hack to phased experiment worth pursuing on Discord. Broadened thesis — community gating is real pain. 95% gross margin on Discord/TG. Start with bot, prove retention lift, then expand. WhatsApp is a margin trap — avoid until unit economics proven elsewhere.
→ 95% gross margin on Discord/TG community gating. WhatsApp API is a 30–45x COGS margin trap.
Founder: Eric San
Agent Creator Directory
8 Feb 2026
Directory/community for AI agent builders — assessed as standalone product
Competitors 100+ directories
Dog-food None
Ceiling $10–35K/mo (1–2 yr)
Best comp cursor.directory $35K/mo
Death Opp cost — time vs Sourcy/Elo
SKIP — ABSORB INTO AGENT ELO
Don’t build a standalone directory. 100+ AI directories exist. Space shows 2000s web directory crash patterns. No dog-food signal. Builder profiles emerge naturally inside Agent Elo when builders submit agents for ranking. The directory is a feature, not a product.
→ 100+ AI directories exist, mirroring the 2000s web directory crash. A directory is a feature, not a product.
Founder: Eric San (+ David Li, Subvert Johnny)
scout
The Filipino Food Data Gap
15 Feb 2026
3 weekend-buildable opportunities across 7 Asian markets — hypothesis, not validated market assessment
Scanned 5 categories × 7 markets
Apps Analyzed 1,050+
Weekend-Ready 3
Top Pick Filipino AI Calorie Tracker
Revenue $0 (TestFlight only)
SCOUT — HYPOTHESIS
Filipino AI Calorie Tracker is the #1 pick. 10-year unsolved pain point (Filipinos can’t track sinigang, adobo, kare-kare calories), zero competition at keyword intersection. NubCal (35K reviews, Thailand) proves a SEA calorie tracker *can* work — but NubCal’s traction is NubCal’s, not Kain’s. Prototype on TestFlight, zero revenue, no deep market research yet. Needs /deepmarketresearch to validate TAM, unit economics, and monetization path.
→ Cross-market arbitrage (what succeeds in JP/KR/TW) is the right starting point for App Store gap detection.
Scout: Eric San
scout
iOS Scout — Hotel & Occasion Apps
15 Feb 2026
Occasion-based hotel booking apps across HK, TW, JP, KR, SG, MY — cross-market analysis + demand validation
Scanned 2 categories × 6 markets
Cross-Market 6 mega-apps (67K–1.15M reviews)
Demand Signals Zero
Weekend-Ready 0
ALL SKIP
Blank space ≠ opportunity. Cross-market analysis: 6 mega-scale apps in JP/KR/TW (Yanolja 953K, Jalan 1.15M, STAYFOLIO 9.6K) all chose NOT to curate by occasion. Demand validation: zero Reddit threads, zero Google Trends data, zero App Store autocomplete for “hotel birthday” / “occasion hotel.” The niche Alice spotted is real on Instagram but doesn’t translate to App Store behavior. All 3 candidates failed validation.
→ Blank space ≠ opportunity. When 6 mega-apps chose NOT to curate by occasion, the absence is signal, not gap.
Scout: Eric San
Vibe-coding wearable hardware — solo founder market assessment
TAM US$318M–1.3B (smart rings)
Reach US$8–18M (Y1–3)
GTM Content-native viral
Margin 37–49%
Break-even ~3K units / US$390K
Death LTV ≈ CAC without subscription
B− CONDITIONAL
Conditionally interesting. High ceiling, fragile floor. 10+ gesture ring companies failed (Nod, Logbar, Ringly, Motiv). Vibe-coding angle is new but hardware supply chain risk is real for solo founder. No recurring revenue = structurally fragile unit economics.
→ 10+ gesture ring companies failed. Hardware supply chain risk is real for solo founders regardless of angle.
Founder: Eric San
stale
Phone Farm SDR Play
22 Feb 2026
When AI identity matters more than AI intelligence. Phone farm economics, on-device model benchmarks, and the identity arbitrage thesis.
Hardware $100 used Android
Value $1,500/mo SDR equivalent
Ceiling 50-phone fleet
Death Platform detection + Legal
SPECULATIVE — $300 PoC
Edge: structural (identity arbitrage). Passes platform bot detection that cloud IPs fail. Institutions can't replicate due to ToS liability. Requires $300 PoC (3 phones) before scaling. Primary kill vectors are platform detection arms race and CAN-SPAM legal compliance.
→ WhatsApp banned chatbots Jan 2026; LinkedIn has active CFAA litigation. Platform risk materializes fast.
Analyst: Eric San
threads strategy
我和我家的90後 — Threads Launch Strategy
25 Feb 2026
Evidence-based Threads strategy for a parent-child podcast launch. 10 ready-to-post drafts mined from full EP1 transcript + viral benchmarks.
Posts 12 ready-to-post
Benchmarks 5 scraped
Top benchmark 5.5K likes
Risk Zero existing followers
READY — LAUNCH POSTS
Mom is the strongest Threads asset. Her quotes are more quotable than any marketing copy. Posts modeled on actual podcast-launch benchmarks (not confession posts). Critical assessment broke 4/7 assumptions from v1 and rebuilt on correct format: story + mom quote + soft pointer.
→ Confession posts drive empathy, not podcast listens. Correct format: behind-the-mic stories + wisdom extraction + reply bait.
Prepared for Eugene Chan
deep investment
Nvidia & The CapEx Supercycle
28 Feb 2026
Q4 FY26 earnings analysis, hyperscaler CapEx, and why retail agentic quant funds should avoid NVDA equity.
Q4 FY26 US$68.1B Rev
2026 CapEx US$600B+
Edge Zero (Equity)
Arbitrage Secondhand Nodes
NO EDGE — DO NOT TRADE
The business is phenomenal, the edge is zero. Trading NVDA equity is a negative-EV game for a retail-scale quant fund. Every institution tracks it. Capitalize on the $600B CapEx downstream by buying depreciating secondhand inference compute, where institutional capital cannot play.
→ No edge = no investment. Capitalize on the CapEx supercycle downstream via secondhand physical compute, not public equities.
Analyst: Eric San
learn
Short Drama Screenplay Skill
5 Mar 2026
Reverse-engineering 短剧剧本工业化生产 — can Donna master industrial AI script production?
Donna Mastered
Eric Optional
Time <5 min setup
Blocker Taste gap (10%)
DONNA MASTERS
Donna can reproduce and execute on command. The 7-step workflow, 4-layer villain system, hook engineering, and bilingual format are all internalized. Eric reviews output for taste. Install in Claude Code for persistent 60-episode projects.
→ Donna masters the craft mechanics. Eric decides what to make and does a taste pass on dialogue.
Analyst: Eric San