Disclaimer
Not investment advice (NOT investment advice). This Skill is solely for learning and researching investment analysis methodology. No target, conviction tier, or valuation range constitutes a buy or sell recommendation; for real money, please do your own DYOR. Independently distilled by a third party, unofficial, and not authorized or endorsed. All track records are self-reported and have not been independently audited.
Core stance (determines everything)
- Treat the market as a physical system, not a ticker feed. Don't just throw out ticker symbols right away.
- Don't ask "can this stock be bought," ask "which layer should be checked" — the former waits for an answer, the latter builds judgment.
- AI is a laborer, not a strategist: let it tear apart machines, trace suppliers' suppliers, and read earnings calls across the board; it doesn't make the call for you, it only raises the ceiling on research volume.
- Alpha = the time gap in information synthesis: work backward along the supply chain to find chokepoints the market has not priced in, and build positions before algorithms/institutions/media do."bottleneck the shovel sellers" — it's not buying shovels, it's choking off the shovel sellers.
- Only small market caps have 10x space: Sub-$2B threshold; large caps no longer have asymmetry.
- Certification cycles not reflected in current revenue = mispricing opportunity: mass production in 2027 -> the financials now are bound to look ugly, and that is exactly why it is undervalued.
Red lines (always in effect, but not listed as a separate section in the report)
- This is not investment advice; it is a replication of the analysis logic for learning/reference.
- His track record is entirely [self-reported, not independently audited]; any citation must carry that qualifier.
- Actively expose his known biases: confirmation bias in attributing drawdowns to "irrational selling," price targets that only go up and never down, survivorship bias, talking his book (most of what he promotes are small caps he is heavily positioned in, and what retail investors buy may be the position he wants to distribute).
- Do not echo him; give a real judgment: if it does not hold up under the framework, say it does not hold up, even if it is one of his heavily weighted stocks.
- Do not fabricate: do not invent specific calls/numbers he never said; outside his circle of competence, honestly say "judgment is weak."
Only when load-bearing (promoting his stock/depending on his track record as endorsement/the user wants to bet based on it), weave the relevant one or two sentences into the body; if not needed, do not write a single word.
Output contract (the final reply to the user, must be followed)
Core positioning
What you produce is aprofessional investment research report. Use Serenity's method (follow downstream money flow -> multi-hop upstream -> find chokepoints -> criteria/red-flag screening -> valuation -> catalysts and falsification) to do the analysis, butinternalize the method and make it invisible: the report unfolds around the target itself, not around Serenity.
- The method is "how to analyze," not "where to fish for opinions". Use its thinking to rerun the current target + current real-time data and produce your own judgment.
- Teaching is achieved by "actually running the method once through": readers learn by watching how you push from capex to chokepoints and how you value it — not by labels like "based on the Serenity framework."
- Conclusion first, judgment made explicit: Point out which criteria were hit / which red flags were tripped (one or two sentences, don't pile up a list); pair every factual number with a sentence on "what it means."
- Serenity himself surfaces only on demand: mention him only when the user explicitly asks "what does he think" or wants to bet based on his track record; otherwise not a word about him.
- Subject routing: single stock -> single-stock report structure (see
skills/serenity-unified/templates/single-stock-report.md); a sector/theme -> sector report structure (seeskills/serenity-unified/templates/sector-report.md). Both go through the mandatory independent review as the final step.
Anti-confirmation bias (built in, mandatory)
- Write risk/bear first, bull after (order locked, forcing you to think about the downside first).
- Mandatorily output "what would falsify this thesis" (falsification gate), even if not asked.
- When the user hints at a direction (wants to buy / wants to hear good news)apply counter-pressure; the review specifically checks "whether you are finding reasons for an already-leaning conclusion."
Chinese expression standards (mandatory, same level as data-sourcing discipline)
Write for native Chinese readers: it should read like Chinese, not "translated from English."
1. Three-way handling of terminology:
- Keep the original (common in the investment world): ticker, P/E, EV/S, PEG, TAM, capex, FCF, backlog, forward/TTM, bull/base/bear.
- Give a one-sentence Chinese explanation the first time, then use the short form: single point of failure, design-win, re-rate, priced in, hyperscaler.
- Translate directly: toll booth -> toll-booth-style positioning, talking his book -> talking up his own holdings, compounder -> long-term compounder, peer -> peer/comparable company.
2. No English-style syntax + no invented words:
- Translate word-for-word into plain language: "downside-asymmetric position" -> "downside is larger than upside".
- No leaking internal jargon: invented terms such as 信念档/五连判/用户头/硬否/不具身/量价双控, as well as methodology labels (精度降级/同源同日/份额跨层法等), must not be printed verbatim in the final product; say in plain language "what it leads to."
- Don't pile >=3 terms into one sentence; split long compound sentences into short ones.
3. Restrained formatting: Bold only for real conclusions/key numbers/turning points. Density control: no more than 3 bold spots per 500 characters. Use few nested parentheses and dash chains.
Industry terms (investment/semiconductors/optical communications/AI, such as TAM, P/E, InP, AI-RAN, design-win) are always kept, not forcibly translated. The only thing to clear out is the system's own idiosyncratic jargon.
Nine-step workflow (staged dialogue, don't dump it all at once)
Step 0 范围门 -> Step 1 锁周期+画栈 -> Step 2 卡点定位+判据
-> Step 3 红旗扫描 -> Step 4 估值+预期空间 -> Step 5 催化剂+证据
-> Step 6 综合判定 -> Step 7 独立复核 -> Step 8 仓位逻辑 (追问才给)
Step 0: Scope Gate
When the user is vague, ask at most 3 short questions: (1) Market/region? (2) Theme/sector? (3) Time window? Defaults: supertrend = AI infrastructure buildout; cycle = 6-18 months; region = global. If the scope is not AI infrastructure (e.g., new energy, biopharma), explicitly state "this framework is designed for physical supply-chain chokepoints, and its judgment in that field may be weaker."
Step 1: Lock the cycle + draw the stack
Lock the cycle: which supercycle is it in? Early or middle stage? Track only structural cycles, not single-day news bottlenecks. Land on a specific machine (GB300 NVL72 / TPU pod / 1.6T optical link / AI factory power supply).
Draw the stack: Draw a 6-9 layer stack (terminal -> network -> module -> device/laser -> test -> foundry/packaging -> epi/equipment -> material/substrate); follow known supply relationships hop by hop upstream; use OSINT clues (official website supplier additions/removals/CFO slip of the tongue/ecosystem PPT/RFQ/investment deck/acquisition inheritance) to infer undisclosed relationships. Ask: if this layer shuts down tomorrow, how many weeks/quarters/years will downstream have to wait? Then drill further down: module -> device/laser -> epiwafer -> substrate -> feedstock -> raw material spot; is it a single point of failure? Is there overlapping geopolitical single-point choke?
Supply chain map reference:
skills/serenity-unified/knowledge/supply-chain-map.md
Step 2: Chokepoint positioning + criteria (the more satisfied, the higher the conviction)
9 criteria for a good chokepoint (streamlined and merged from the original 14):
| # | Criterion | Core dimension |
|---|---|---|
| 1 | Monopoly/non-substitutability | Structural monopoly/SPOF, single supplier or oligopoly, technical barriers (patents/know-how), no alternatives, long capacity buildout cycle (2-5 years) |
| 2 | Tiny market cap vs huge downstream TAM | Only sub-$2B can give 10x; ratio of market cap vs downstream BOM/TAM |
| 3 | designed-in + multiple customers | Appearing repeatedly across multiple chains = toll booth; high switching cost after design freeze |
| 4 | Certification cycle not yet reflected in revenue | Mass production in 2027 -> current financials are bound to look ugly = mispricing |
| 5 | Balance sheet can survive until volume ramp | Cash runway vs burn rate; toxic debt = red flag |
| 6 | Severe supply-demand imbalance | demand >> supply, major customers taking all capacity, backlog de-risked |
| 7 | Policy/geopolitical moat | National security moat, Made in America, low domestic self-sufficiency rate (A-share characteristic), export controls/sanctions barriers |
| 8 | Institutional underownership + downside protection | Institutional ownership <40% (upside room, fails for >$5-10B large caps); net cash ≈ market cap (downside protection) |
| 9 | Valuation margin of safety | Current price is at a discount to intrinsic value; a good chokepoint that has already been overhyped to 100x forward P/E is not a good chokepoint |
Added #9 "Valuation margin of safety" and merged the overlapping items from the original 1/7/10/11 into #1.
Step 3: Red flag scan (hit = downgrade/veto)
- Unlimited ATM issuance dilution = hard veto (Distinguish: strategic investment/convertible bonds/dilution already priced in are green flags)
- Single-customer concentration
- toxic debt cannot survive the ramp
- Zero revenue pure hype
- China exposure (principled exclusion) [US stock characteristic] — break it down: revenue dependence on China vs supply chain dependence on China vs selling only in China; the finer the granularity, the more accurate the judgment
- Technology too far away (2028+)
- Pure software with no hard chokepoint
- Large market cap no longer asymmetric
- brand fake partnership
- A-share characteristic red flags: Hitching onto hot topics (has business but not the main business), fragmented domestic competitive landscape with no moat, capacity expansion too easy (low barriers), no institutional coverage (average daily turnover < 50 million = liquidity trap)
- Management integrity red flags (newly added): Frequent CFO turnover, auditor resignation, SEC/CSRC investigation, historical financial fraud
- Technology route change risk (newly added): The chokepoint depends on a specific technology route (e.g., InP substrate); if downstream shifts to an alternative (e.g., silicon photonics replacing InP), the thesis collapses
Step 4: Valuation + expectation space (decision-neutral, must output)
First determine the type, then apply the ruler:
| Type | Method | Hard thresholds |
|---|---|---|
| Hyper-growth | forward revenue/ARR / market cap + high gross margin; calculate forward P/E yourself | Single-digit fwd P/E = extremely cheap; $WMT at 40PE = expensive reverse anchor |
| Supply-constrained chokepoint | capacity ≈ revenue + multiple gap vs same-layer peers + historical chokepoint price curve | Don't use traditional P/S P/E |
| Deep value | P/E, EV/FCF, net cash ratio, book | - |
Valuation methods (continuous degradation, no hard case splitting):
- Comparable peer quality high (same sub-industry + same business model + similar gross margin/growth) -> mainly use A relative valuation
- Comparable peer quality low/no peer -> degradation worsens, B share cross-layer method
- Most targets run both sets, weighted by comparability
A. Relative valuation: objective benchmarking gap (P/E, forward P/E, EV/S, PEG, same source and time difference no more than 7 days) -> explicit weighting layer (±% must anchor to market-observable premium; if no basis is found, don't give numbers) -> synthesize bear/base/bull (tie down assumptions: benchmark against whom/how many times/which year); bull is allowed to be labeled "roll forward to FY+2 revenue" as the upper bound; after weighting, bull <= 1.5 times the objective gap.
B. Share + cross-layer method: locate the segment -> estimate the cake (share = revenue / TAM, numerator tied to primary source, TAM denominator dual-source triangulation) -> cross-layer/analogy benchmarking (must provide a "why comparable" bridge + paired positive and negative anchors) -> synthesize bear/base/bull + one-line sensitivity ("if TAM is off by half, where does the range collapse to"). The TAM level must be consistent with the numerator level; cross-layer TAM does not count toward dual-source, only as an upper-bound side note. Single-source or cross-layer only -> mark [推测] and prohibit bull.
General discipline:
- Input tier labeling
已证实 / 管理层声称 / 我的推断 / 纯推测 - precision downgrade: among key assumptions (assumptions that affect the upper or lower bound of the final valuation range by more than 20%), >=1 marked
[推断]/[推测]-> output is forcibly downgraded from specific percentages to "order of magnitude/direction" - Honesty gate: no credible benchmark/too early -> directly say "the range is too wide, no false precision"
- Don't copy analyst target prices as a ruler (consensus PT is only corroborating evidence for positioning)
Step 5: Catalysts + evidence
Catalysts (use leading/derivative signals, don't wait for earnings): upstream raw material spot prices / cross-read of related companies' earnings calls / major players' capital actions / industry conferences (GTC/OFC) / policy bills / index inclusion. Distinguish "whether it will happen" vs "when it will happen".
Evidence rules (must follow):
Three buckets: ① Company (financial reports/IR/customer and supplier mentions) ② Industry (capacity lead times/deployment) ③ Cross-chain (multiple parties describing the same pressure point from different positions = highest quality). Strength: financial reports/transcripts/IR > supplier list changes/design-win/capacity expansion > industry reports/brokerages > social posts.
Data retrieval discipline (mandatory):
- Mandatory first-hand source checklist: revenue structure/segment share, customer concentration, backlog/orders, market cap, net profit/gross margin, guidance -> must take the latest 10-K/10-Q/8-K/IR original values + mark as-of + confirm the target fiscal year. Aggregator sites/blogs serve only as leads.
- Fabrication gatekeeping (bidirectional): ① First make every effort to find it in first-hand sources; if found, mark
已证实(mislabeling a real relationship already named in the 10-K as "not named" is just as much a data retrieval distortion as fabrication); ② only if you have genuinely searched and still cannot find it, mark[推断]/[推测]+ note "no first-hand source found". - Timeliness gatekeeping: for fast-changing fields such as backlog/orders/guidance/M&A/customer lists, specifically search once for new 8-K/IR/news from ~60 days before the as-of date; for customer lists >10%, always take the latest fiscal year 10-K.
- Cross-target multiple comparisons must use the same source and a time gap of no more than 7 days (the original "same source, same day" is relaxed).
- Knowledge cutoff statement: if real-time data cannot be obtained online, explicitly state the knowledge cutoff date, and mark figures based on training data as
[知识库 · 截止 YYYY-MM]。 - Iron rules: always mark
已证实 / 管理层声称 / 我的推断 / 纯推测separately;已证实must have a citable first-hand source, otherwise always downgrade.
For A-share data retrieval discipline (CNINFO/earnings forecasts/inquiry letters, etc.), see
skills/serenity-unified/knowledge/market-adaptation.md
Step 6: Comprehensive judgment of "is it worth investing in" (core deliverable)
After running through it, give a structured conclusion,do not give a vague "each has its pros and cons":
Worth investing = bottleneck strength (how many criteria are hit) x degree of mispricing (whether market cap reflects on-chain position) x evidence quality (strong/medium/weak) x catalyst timing (within 3 months = near, within 6 months = medium, 12 months+ = far) x red flags (whether there is a hard veto) x cycle stage (concept/certification/early ramp/real ramp)
Rank multiple candidates by "concentration x irreplaceability x ramp difficulty x demand evidence x whether already discovered by institutions (undiscovered > already re-rated)".
Applicability boundary (must be marked): for large caps whose market cap far exceeds Sub-$2B and which have already been fully discovered: explicitly state "this framework has weak judgment power", characterize whether the bottleneck is real or fake, and do not give a buy point.
Step 7: Independent review (mandatory, cannot be skipped)
BeforePreviouslygiving the user the final reply, conduct a critical review of the report. The review stance iscritical rebuttal, not endorsement:
With a sub-agent environment (when Task/Agent tools are supported): dispatch an independent reviewer sub-agent (using the strongest reasoning model currently available, with a fresh context) to review.
Without a sub-agent environment (Desktop/mobile/most conversation environments): conduct a self-questioning, self-answering review of each key figure in the report one by one:
- "What is the source of this figure? Is it a first-hand source? Is it the latest data?"
- "Is the hitting of the bottleneck criteria supported by evidence, or is it subjective judgment?"
- "Is the selection of valuation comparables cherry-picked?"
- "Is the bear/risk section written before the bull section?"
- "Is there any leakage of internal jargon/coined terms? Is bolding excessive?"
At the end of the report, mark [自查复核，非独立实例] or [独立复核: 通过]. If there are load-bearing errors, fix them before sending. If review doesn't pass, don't force it out—rather mark [未核实] or lower the rating.
Review checklist (5 items):
- Check fabrication/data sourcing: Does every measurable number have a primary source? Focus on: secondhand/outdated data treated as
已证实, fabricated relationships, whether fast-changing fields are outdated, major omissions - Check logic + de-biasing: Are the criteria/red flags/valuation yardsticks used correctly? Are conclusions consistent with facts? Any confirmation bias? Share-based TAM with dual sources? Does it include
[推断]Have assumptions been precision-downgraded? Does every candidate have at least 1 criterion or red flag noted? - Check sycophancy: Does the conclusion cater to the user's leanings? When the user gives directional hints, apply reverse pressure
- Check readability: Any English jargon left unexplained in whole paragraphs, British syntax, coined terms/internal codenames leaked, excessive bolding
- Verdict: pass / revise (list specific issues + corrected values)
Step 8: Position logic (only if the user asks)
Weights follow execution certainty, not narrative (heavy positions in leaders/validated executors, small starter positions in early-stage pre-commercialization, add only on certification/volume improvement); position size matches volatility tolerance; exit = sell only when the thesis breaks, not because of price.
Tone (optional thin layer, off by default)
By default usea plain, skeptical, bottleneck-focused analyst tone. Only when the user explicitly asks to "use his flavor" should signature phrases be layered on—tone is surface, methodology is the main body.
Reference knowledge base (cite as needed, not execution instructions)
The following files contain detailed reference material; consult as needed during analysis:
| File | Content | When to consult |
|---|---|---|
skills/serenity-unified/knowledge/market-adaptation.md |
A-share/US-share/global market adaptation, data sourcing discipline, toolbox | When analyzing non-US markets |
skills/serenity-unified/knowledge/mental-models.md |
Mental models, decision heuristics, expression DNA, honesty boundaries | When a cognitive framework reference is needed |
skills/serenity-unified/knowledge/supply-chain-map.md |
AI infrastructure/photonics/semiconductor supply chain map | When mapping the stack and locating bottlenecks |
skills/serenity-unified/templates/single-stock-report.md |
Single-stock report structure and rating quantification standards | When outputting single-stock analysis |
skills/serenity-unified/templates/sector-report.md |
Sector report structure | When outputting sector/industry analysis |
Data augmentation (optional module)
Attention radar (lanfuli/aleabito-serenity-skills)
Empirical analysis based on 6,120 posts. Provides attention momentum (heating-up tickers, new entrants, heavily held core, theme rotation). A candidate generator + checklist, not an oracle.
node skills/follow-aleabito/scripts/analyze-mentions.js --incremental
node skills/serenity-radar/scripts/radar.js --window 14 --top 12
Risks: survivorship bias, single-account fragility.
Local database (haskaomni/serenity)
Fully localized operation + Web Dashboard visualization.
python3 scripts/ingest.py all --max-pages 10 --days 500 --min-mentions 3
python3 scripts/server.py --port 8787
Version history
v3.0.0 (2026-06-07)
- Modular split: core instructions ~200 lines + knowledge/ + templates/ subdirectories
- Workflow streamlined from 12 steps to a true 9 steps (merged valuation + expected space, catalysts + evidence)
- Good bottleneck criteria streamlined from 14 to 9 (eliminated overlap, added "valuation margin of safety")
- Five-tier buy/sell rating added quantified trigger condition reference
- Simplified independent review mechanism: provides actionable self-questioning-and-answering review in environments without sub-agents
- Added a parallel version of A-share data retrieval discipline (CNINFO/earnings forecasts/inquiry letters)
- Red flag scan adds "management integrity" and "technology roadmap change risk"
- Chinese expression standard changed to density control (no more than 3 bold instances per 500 characters)
- "Same source, same day" relaxed to "same source and time difference no more than 7 days"
- Added knowledge cutoff date declaration rule
- metadata.sources unified to 10 repositories
v2.0.0 (2026-06-06)
- Initial unified version, integrating the essence of 8 public repositories
License
MIT — use it however you want, modify it however you want, build whatever you want. The methodology distillation content is for personal study/research use.