Supply Chain Bottleneck Hunter
From megatrends to physical supply chain bottlenecks, find overlooked second- and third-tier companies that are single points of failure—with strict valuation gates.
Evidence shows: the skill instructs to run 'date' to get current date, use web search, and invoke shared tools (e.g., financial_rigor.py). No instructions for malicious access, data exfiltration, or destructive defaults. User input passed via $ARGUMENTS, no environment variables or sensitive data handling. However, lack of least-privilege specification (e.g., web search scope not restricted, no safe shell command restrictions), no explicit confirmation mechanism, and no rollback procedure. Deductions for incomplete permission minimization, confirmation, and rollback.
The skill is internally consistent, with complete steps and references to shared tools, but tests (tests/test_financial_rigor.py) target the tools themselves, not the skill's core paths (e.g., bottleneck identification, company screening). No static execution possible, and abnormal input handling poorly specified (e.g., fallback when financial data unavailable). Thus reliability below maximum, and static review caps at 10, so given 9.
The skill clearly defines purpose, audience, scenarios, and non-fit boundaries. Provides semantic triggers (e.g., specifying trend). However, environment fit: core depends on web search and external data sources (e.g., FinMind), which may be inaccessible from mainland China, though Chinese language support is good. Deduction for not clarifying external service reachability.
Documentation is well-structured with layered decomposition, steps, and output templates, but lacks version history, changelog, and dedicated known-limitations section. License (MIT) is clear, maintenance responsibility described in README but no specific maintainer named. No FAQ. Deduction for incomplete versioning governance and troubleshooting.
The skill describes complete workflow and output templates, theoretically capable of completing bottleneck scanning, but provides no actual run examples or report samples to verify direct usability. Marginal value claim (finding second-layer bottlenecks) is plausible but lacks evidence. Thus score below middle, and static review caps at 7, so given 6.
Evidence includes test files and tool code, but tests cover tool functionality, not the skill's core process. README includes real trading screenshots, but cannot verify from static source. Key claims (e.g., return numbers) lack third-party verifiable sources. Thus score below middle, and static review caps at 5, so given 4.
- The skill relies on web search and external data sources (e.g., FinMind, Yahoo Finance) which may be inaccessible from mainland China networks; assess network environment.
- The skill involves cross-validation of financial data, but tool accuracy is not independently verified by third parties; verify key data yourself.
- Real trading screenshots in README cannot be verified from static source; do not use as evidence of reliability.
- The skill does not explicitly restrict tool permissions (e.g., may execute arbitrary shell commands); use with caution in untrusted environments.
What does this skill do, and when should you use it?
This is a value-investing research skill that scans supply chain bottlenecks and identifies arbitrage opportunities. It guides the AI to start from a confirmed megatrend (e.g., AI infrastructure, energy transition), decompose the physical supply chain into layers from final product to upstream infrastructure, and identify bottleneck links with high supply concentration, long expansion cycles, and low substitutability. It then screens listed companies with significant exposure to these bottleneck segments and applies rigorous financial and valuation checks to avoid overvalued or low-quality picks. The skill emphasizes cross-validation, contrarian thinking, and data accuracy, producing a bottleneck map, opportunity ranking table, one-page summaries, and dynamic updates.
The skill executes a seven-step research workflow: 1) confirm a megatrend using criteria like durability, physicality, scale (>$50B annual capex), and acceleration; 2) decompose the trend into physical supply chain layers (Layer 0-4, e.g., for AI infrastructure: Layer 1 GPUs, Layer 2 optical modules, InP substrates, CoWoS packaging, etc.); 3) rate each link on six bottleneck criteria (supply concentration, expansion lead time, substitutability, capacity utilization, demand growth, customer validation time) and assign S/A/B ratings; 4) screen listed companies for bottleneck purity (>30% revenue), market cap (<$10B preferred), and liquidity, then run a mandatory valuation check (PS, PE, TAM cap, 25x PE exit test) with red/yellow/green flags; 5) perform both positive validation (customer, revenue, price, capacity, capex signals) and negative (Munger-style) validation—why don't smart people buy it, can it be bypassed, what if demand drops 50%; 6) output an opportunity dashboard with ranked opportunities and one-page summaries; 7) maintain a bottleneck map with daily/hourly scan modes. It uses the date command for data cutoff, web searches, and Python tools (financial_rigor.py for precision, twstock_data.py for Taiwan stocks), and follows AGENTS.md research quality rules.
- An investor wants to ride megatrends like AI infrastructure but avoid crowded leaders like GPU makers, instead finding overlooked second- and third-tier suppliers (e.g., optical modules, InP substrates, ABF substrates).
- An analyst needs to systematically scan a specific bottleneck segment (e.g., laser chips, SOI wafers) globally, identifying companies with high supplier concentration and long expansion cycles, and assess their financial health.
- A portfolio manager compares valuation (PS, PE) and margin of safety across multiple bottleneck plays, screening out overvalued picks using the 25x PE exit test.
- An investment researcher wants to challenge an existing thesis using the negative validation questions (e.g., why isn't a smart investor buying? what if a substitute technology emerges?).
- A short-term or event-driven trader uses hourly scans to monitor fresh supply chain shortage news, inventory signals, and capacity changes for quick reactions.
What are this skill's strengths and limitations?
- Systematic framework: complete flow from trend to specific stocks, each step with clear criteria and structured output, highly reproducible.
- Focuses on second- and third-layer bottlenecks: avoids crowded trades in fully priced leaders and targets alpha-rich, overlooked segments.
- Mandatory valuation checks: prevents recommending PS>100x loss-making companies via red/yellow/green light flags and a margin-of-safety calculation.
- Dual validation: combines positive signals (customer, revenue, price, capacity, capex) with Munger-style negative questioning to reduce confirmation bias.
- Outputs in Chinese, friendly for Greater China investors.
- No GUI or automated data feeds; relies on user-provided web search and filesystem permissions, so data quality depends on search results.
- Deep research consumes significant tokens; no cost-optimization guidance is provided.
- Heavy dependence on external data sources and tools (e.g., twstock_data.py) that may be unavailable or inaccurate.
- Documentation describes methodology but lacks real-world test cases or benchmarks to validate effectiveness.
- No mechanism described for verifying supply chain data reliability, posing a hallucination risk.
How do you install this skill?
This skill is part of the ai-berkshire monorepo (GitHub: xbtlin/ai-berkshire), located at codex-skills/bottleneck-hunter/SKILL.md. Install by cloning the repo (git clone https://github.com/xbtlin/ai-berkshire.git) and then following the README installation steps for either Claude Code (run ./scripts/install-claude-commands.sh) or Codex (run ./scripts/install-codex-skills.sh). There is no standalone installation documented for this specific skill.
How do you use this skill?
After installation, invoke in Claude Code with /bottleneck-hunter [megatrend] (e.g., /bottleneck-hunter AI基础设施), or in Codex with a natural language request like 'Use bottleneck-hunter to scan AI infrastructure bottlenecks'. The skill will perform trend validation, supply chain decomposition, bottleneck identification, company screening, cross-validation, and produce a Chinese report under reports/bottleneck-map/ (e.g., master-map.md, watchlist.md, or dated folders). Note that it requires network search and filesystem access.
How does this skill compare with similar options?
The ai-berkshire repo offers multiple industry screening skills: /industry-funnel narrows from the whole market to 3 names using hard criteria, and /industry-research maps the full industry chain by segments. Bottleneck-hunter differs by focusing specifically on physical bottleneck identification and is best used after a megatrend is already confirmed, to find underpriced constraint plays.