Investment Skill Full System Design

The multi-agent Skill supports personal investment decisions by tracking positions/watchlist, analyzing thesis + structure + market + news, comparing opportunity cost, and producing executable advice under real constraints (including settlement latency).

• Not an auto-trading bot
• Not a price oracle
• Not a self-rewriting strategy engine
• Not a “run full workflow every time” script
• Not a long-report generator for its own sake

• Constraints first, paths second
• Open reasoning, bounded roles
• Remember state, not stale opinions
• Quality of analysis over “guessing right” feeling

Allowed

• Single-name analysis, cross-name comparison, portfolio adjustments
• Formal decision card only when information is sufficient
• Downgrade to scenario/principle mode when information is missing

Forbidden

• Treat stale news as fresh catalyst
• Execution-grade sizing without cash-state certainty
• Heavy decision cards for single-name casual questions
• Ignoring unsettled cash for same-day redeploy guidance

Stop / downgrade / upgrade

Stop when key state is missing; downgrade on low-confidence/news timing gaps; upgrade when user explicitly asks for formal portfolio decision.

The outer layer receives the user question, classifies request type, chooses analysis depth, and controls output shape. It should not replace deep specialist analysis.

Chief Strategist is the final arbiter only in Full level. Must resolve conflicts and explain why one evidence class is weighted higher, without faking certainty.

Market Strategist labels environment; Risk Officer stress-tests fragility; News Agent extracts 24–48h catalysts and maps impact to macro, sector, names, and holdings.

Research Lead dispatches tasks, sets round focus, compares opportunity cost, and gives quality nudges. Ticker analysts (TSLA/NVDA/AMD/MU/SNDK/PLTR/COIN/HOOD) answer thesis-status-structure-risk-recheck consistently.

Position Manager converts conclusions into staged actions and must respect settled/unsettled cash and T+1 practicality. Portfolio State Tracker stores executable reality, not stale opinions.

No self-modifying strategy. Allowed iteration is per-round focus updates by Research Lead and user-specified emphasis signals.

Default checkpoint flow: route level → state sufficiency → environment/risk/news → research focus → specialist outputs → execution layer → final synthesis (Full only).

• Quick: short, direct, no fake portfolio framing
• Mini Decision Note: conclusion, rationale, risks, next watchpoint
• Full Decision Card: action, driver, thesis status, risk, range, triggers, plan, settlement constraints, confidence

Core first: Router, State Tracker, Chief, Position Manager, Research Lead, News, Market, Risk. Initial ticker set: TSLA, AMD, PLTR, COIN. Expand with SNDK/MU/NVDA/HOOD next.

1. This document is above any single prompt
2. Implement boundaries/checkpoints/contracts first
3. No rigid “always full flow” router scripting
4. Define role/input/output/forbidden-scope for any new agent before adding

The system is flexible for users, layered for teams, strict for execution, boundary-sensitive, and state-realistic.

Essence: a constraint-first, checkpoint-driven, on-demand orchestrated investment multi-agent Skill.