Drawdown Decision Checklist

Use this reference after reading SKILL.md when the user asks whether to endure, reduce, exit, or reframe an AI hardware / physical AI selloff.

Required Inputs

Collect what is available, and mark the rest unknown:

Input	Why it matters
Current date/session	"today" and "latest" answers need live evidence.
Current price, prior close, intraday range, volume	Distinguishes normal volatility from disorderly selloff.
Sector basket/index movement	Separates company-specific issue from factor rotation.
News/catalyst	Determines whether there is a new fact or only price pressure.
User holdings and weights	Determines concentration risk.
User cost basis and max tolerable drawdown	Determines whether the position is too large even if thesis is intact.
Time horizon	Same-day panic, 1-2 week stabilization, next-earnings validation, and multi-quarter thesis are different decisions.
Overnight US large caps	AI hardware often follows NVIDIA, AMD, Micron, Broadcom, Meta, Microsoft, Google, Amazon, Nasdaq and SOX.
Japan/Korea/Taiwan semiconductor session	A/H hardware names often react to Samsung, SK hynix, Tokyo Electron, Advantest, TSMC and regional chip indices.
Core narrative shock	Short-term messages such as "excess compute", "capex cut", "export control", "CPO delay", or "HBM oversupply" can hit the story before earnings prove or disprove it.

If user-specific inputs are missing, say so and give a generic framework.

Green / Yellow / Red Triage

Triage	Meaning	Evidence pattern	Response framing
Green	Thesis intact, price/crowding damage only.	No order/capex/guidance/margin break; peers and factor basket sell off together; primary sources still support demand and bottleneck.	Do not treat the day as thesis exit; continue only if position risk is acceptable.
Yellow	Thesis not disproven, but risk budget or valuation is wrong.	Large prior run, crowded basket, insider/major-holder sales, analyst downgrades, stretched valuation, missing financial proof, or user cannot tolerate another leg down.	De-risking or rebalancing can be rational without calling the thesis false.
Red	Thesis damage or risk breach.	Capex cut, order cancellation, customer loss, guidance miss, margin failure, accounting/regulatory issue, export controls, demand delay, capacity relief, or substitute architecture.	Treat as thesis review/exit condition; require fresh primary-source evidence before re-entering the old thesis.

AI Hardware Evidence Gates

Demand and capex

• Hyperscaler/cloud capex: actual capex, capex guidance, RPO, data-center construction plans.
• Model/application demand: paying usage, enterprise adoption, run-rate revenue, inference cost pressure.
• China/local demand: official cloud/AI spending, procurement, export-control impact, domestic substitution.

Overnight lead-lag screen

Use this screen when the user asks about tomorrow or the next session:

Lead signal	Bullish / stabilization read	Bearish / follow-through read
NVIDIA / Broadcom / AMD	Flat-to-up after prior selloff; no new capex/order bad news.	Down materially with negative volume or fresh order/capex concern.
Micron / HBM / memory	Stops falling; Samsung/SK hynix stabilize in Asia.	Memory leaders keep falling; oversupply narrative spreads.
Meta	Rises while AI suppliers stabilize: monetization story absorbed.	Meta rises but suppliers keep falling: market reads "customer becomes competitor" or "spare compute" as supplier risk.
CoreWeave / Nebius / neoclouds	Rebound or management/source rebuttal supports demand.	Continued sharp decline: rental-capacity business model under pressure.
Nasdaq / SOX	SOX stabilizes relative to Nasdaq.	SOX underperforms Nasdaq by a wide margin.
Japan/Korea/Taiwan semis	Samsung/SK hynix/TSMC/TEL/Advantest stabilize before A/H open.	Regional chip bellwethers gap down again or break prior lows.
Domestic pre-open news	No new exchange risk warnings or regulatory shock; broker notes distinguish narrative shock from capex cuts.	More abnormal-trading warnings, margin pressure, capex-cut rumors, or policy/export-control shock.

Translate the screen into scenarios:

Scenario	Trigger	Domestic AI hardware read
weak_open_continue	US semis and Korea/Japan keep falling.	Expect weak open and pressure on high-beta CPO/PCB/memory/AI server names.
panic_rebound	US semis stabilize, neoclouds stop falling, Korea/Japan stabilize.	Expect rebound attempt, but prefer leaders with earnings/order proof over pure concepts.
sector_rotation	Index stable, AI hardware weak, low-valuation/defensive sectors strong.	Theme is still being de-risked; do not read broad index strength as AI hardware strength.
thesis_damage_follow_through	New primary-source capex/order/guidance cut appears.	Treat as thesis review, not just technical pullback.

Narrative Shock Examples

Do not treat every narrative shock as thesis damage. First decide what part of the story it attacks:

Shock	Story attacked	Hard proof needed before calling thesis damage
Hyperscaler plans to rent excess compute	"AI compute is always supply constrained" and neocloud scarcity economics.	Actual capex cut, GPU order cancellation, utilization drop, rental-price decline, or supplier guidance cut.
CPO or architecture delay	"Optics/CPO bottleneck timing."	Customer roadmap delay, vendor guidance cut, volume-production timeline slip.
HBM/memory price panic	"HBM shortage and pricing power."	Contract price decline, inventory build, customer qualification shift, capacity relief faster than demand.
Major customer insourcing	"Supplier chokepoint and order duration."	Named order reduction, RPO/backlog decline, lost design win, dual-source confirmation.

Orders and conversion

• Signed orders, backlog, RPO, prepayments, long-term agreements, named design wins.
• Shipment growth, utilization, delivery windows, ASP, customer concentration.
• Revenue and gross margin conversion in the latest quarter or half-year.

Bottleneck persistence

• Expansion lead time, qualification cycle, yield/ramp risk, constrained feedstock/materials.
• Customer ability to dual-source, redesign, insource, or switch architecture.
• New capacity or second suppliers that could relieve scarcity.

Financial quality

• Latest revenue, net income/loss, gross margin, operating cash flow, capex, free cash flow.
• YoY/QoQ comparison and whether the AI/robotics segment is big enough to matter.
• Balance sheet, dilution, debt maturity, customer credit, inventory quality.

Portfolio Bucket Rules

Bucket	Use for	Typical response
core_chain	Direct AI capex bottleneck with measurable orders/revenue/margin.	Hold the thesis only while evidence gates stay intact; review valuation and sizing.
measurable_supplier	Public supplier with financials but incomplete AI/robot-specific purity.	Watch orders, segment mix, margins, and customer proof.
satellite_option	Robotaxi, humanoid, edge AI, or optional future chain with catalysts but limited revenue proof.	Keep smaller/risk-aware; do not let concept exposure dominate.
separate_thesis	Stock is not really AI hardware, but may have a different cyclical/value thesis.	Evaluate on its own fundamentals; do not defend it with AI capex logic.
overcrowded	Strong theme but price has outrun evidence or social heat is high.	Reduce risk if user sizing is large; require evidence before increasing conviction.
needs_evidence	Relationship to chain is unclear or source gaps dominate.	Keep as watchlist until primary evidence appears.

Sunk-Cost Rebuy Test

Use this when the user already has a loss and asks whether to endure, wait for a catalyst, or leave. The test is:

The loss has already happened. Do not decide based on getting back to breakeven. Ask: if this position were cash right now, would the user willingly open the same size before the next risk window?

Interpret the answer:

Answer	Meaning	Response framing
Yes, same size	The current thesis and risk budget still match.	Continue only while the evidence gates and stop conditions are explicit.
Yes, but smaller	The thesis may remain, but sizing is wrong.	Separate a smaller catalyst/event lot from the failed or oversized short-term lot.
No	The position is being kept mainly because of sunk cost or hope of breakeven.	De-risking is justified by position risk even if the long-term theme is not disproven.
Unsure or emotionally overloaded	The user cannot judge the thesis cleanly at the current size.	First reduce to a size that permits calm review, then reassess facts.

Apply it before letting a later catalyst override today's failed trade:

• A short-term chase that fails does not automatically become a medium-term thesis lot.
• A medium-term catalyst, such as an IPO, ADR listing, earnings date, product launch, or index event, can justify only the size the user would newly open today.
• The maximum "endure" size is the amount the user can still hold if the next session gaps down another realistic stress amount, such as 5%-8% for high-beta AI hardware names.
• If the user would not newly buy the same size, keep the wording focused on "risk budget mismatch" rather than "the thesis is false."

Close-First Timing Discipline

Use this when a user considers adding, reducing, or switching positions during a volatile session, especially after a gap, a prior one-day surge, a news catalyst, or heavy volume.

Core rule:

On high-disagreement days, morning price action is evidence, not confirmation. Prefer making larger decisions near the close, when the session structure, sector breadth, and cross-market read-through are clearer.

Interpret the intraday phase:

Phase	Default interpretation	Response framing
First 30-60 minutes	Price discovery and disagreement are highest.	Avoid treating a morning spike or dip as final confirmation; only observe or use small risk-defined test lots.
Midday	Better signal only if liquidity, breadth, and leadership stabilize together.	Consider opportunity only when the target reclaims key intraday levels, sector peers stop falling, and invalidation is explicit.
Late session / close	The market reveals whether the move is accumulation, distribution, or failed rebound.	Use the close to decide whether the setup is improving, still broken, or only a trade.
One-way surge into close	Momentum may be real, but chase risk rises before the next session.	Do not treat a vertical close as permission for full-size exposure; require a next-day plan for high-open fade or failed follow-through.

Minimum conditions for a midday exception:

• Overnight/global lead-lag is supportive rather than contradictory.
• Local sector breadth stabilizes, not only one name bouncing.
• The target reclaims meaningful intraday levels such as VWAP, open, prior close, or the morning breakdown area.
• Position size, invalidation level, and same-day/next-day handling are defined before entry.
• The trade remains a small opportunity lot unless close confirmation arrives.

When the stock closes with a one-way surge:

• Do not automatically chase full size into the close.
• If the user buys the extended close, label it an event or momentum lot, not a long-term thesis lot.
• Predefine the next-session response: if it opens high but cannot hold, use the failed follow-through as a risk-reduction signal; if it holds above the prior close with sector confirmation, only then treat it as stronger right-side evidence.
• If the close is strong but the user cannot tolerate an overnight gap down, skipping the entry can be rational even if the theme remains attractive.

Post-Selloff Rebound Confirmation Checklist

Use this after a sharp sector selloff to distinguish a tradable oversold bounce from an early stabilization candidate. Check all three conditions at or near the close:

Check	Pass condition	Failure meaning
No new low	The local AI hardware basket and relevant leaders hold above the prior panic-session low.	Selling pressure is still extending; an intraday bounce has not repaired the structure.
Volume contracts	Turnover/volume declines versus the panic session while prices stop making new lows.	Heavy volume with weak prices suggests distribution, forced selling, or unresolved disagreement.
Rebound holds into the close	The sector and leaders retain a meaningful part of the rebound and avoid a late-session return to the low.	A morning or midday rebound that fades into the close remains an oversold bounce.

Decision rule:

• 0-1 passed: OBSERVE or REDUCE_RISK; treat the move as an oversold bounce or continued price discovery.
• 2 passed: OBSERVE; stabilization is possible but not confirmed. Record the failed condition for the next session.
• 3 passed: TEST_ONLY; label it a rebound-confirmation candidate, not a trend reversal. Require next-session follow-through and global semiconductor confirmation before upgrading further.

Do not substitute one strong stock for sector breadth. Apply the checklist to the relevant basket first, then confirm that the user's holding is not materially weaker than that basket.

Decision Language

Use scenario language instead of orders:

• "This does not yet look like thesis damage because..."
• "It becomes thesis damage if..."
• "The reason to reduce is risk budget/valuation, not because the chain is disproven."
• "Keeping exposure only makes sense if you can tolerate X and will verify Y."
• "This name should be separated from the AI hardware bucket."
• "If this were cash now, only the portion you would willingly reopen belongs to the catalyst bucket; the rest is a failed or oversized risk lot."
• "Morning moves are evidence, not confirmation; the larger decision should wait for close confirmation unless the midday setup is independently confirmed and risk-defined."

Do not say:

• "Buy the dip."
• "Must hold."
• "Sell immediately."
• "It will rebound."
• "The market is wrong" without primary evidence.

Short Output Template

结论：
今天更像 <green/yellow/red + classification>。这不是投资建议。

事实底座：
- 价格/指数：
- 新闻/催化：
- 证据强弱：

持仓分层：
| 标的 | 桶 | 这次跌法说明什么 | 下一验证 |

触发器：
- 继续跟踪：
- 降低风险：
- 重新评估/退出旧 thesis：

下一步：