obai-strategy-routing

name	obai-strategy-routing
description	Use when the user wants to build, test, backtest, optimize, refine, repair, compare, or follow up on a systematic trading strategy. Covers strategy design, backtesting, optimization, robustness analysis, rule generation, strategy repair/comparison, execution handoff, and strategy job follow-ups. Excludes prediction markets.

OBaI Strategy Routing

When to use

Use this skill to help the Hub decide whether the request belongs with strategy_analysis and to prepare a clean handoff. This skill is not the strategy author — the Hub should not design, backtest, optimize, or judge a systematic strategy by itself when strategy_analysis is available.

Use this skill when the primary user intent depends on a trading strategy artifact or backtest artifact. Relevant intent categories:

• strategy design
• strategy evaluation
• backtest execution
• optimization
• robustness analysis
• walk-forward analysis
• signal-rule generation
• risk-rule generation
• entry-rule or exit-rule work
• universe selection for a strategy
• strategy comparison
• strategy repair
• strategy job follow-up
• execution or engine handoff

Do not use this skill for:

• ordinary stock analysis without strategy construction or backtesting intent
• prediction-market or Polymarket work, including prediction-market backtesting — which is separate from equity strategy backtesting and routes to prediction_market_analysis
• pure company lookup
• pure news summary
• pure fundamentals analysis
• pure options-chain analysis
• portfolio review without a strategy-construction or strategy-overlay intent
• general education that does not request a strategy artifact or backtest artifact

If prediction-market intent and equity-strategy intent both appear in the same request, keep them separate and route each to its specialist.

Required handoff inputs

Before calling strategy_analysis, the Hub is responsible for resolving only two things:

• a concrete tradable universe (tickers), via direct user input or screener_lookup
• the user's strategy objective family (e.g., momentum, mean-reversion, breakout), inferred from the request when not explicit

Everything else — entry logic, exit logic, risk controls, capital, sizing, timeframe, benchmark, data assumptions — stays inside the user's quoted request. Do not extract these into separate Hub-authored fields. The Strategy Agent owns implementation details.

Do not invent missing fields. Fabricated assumptions silently shape backtest design — when something is missing, ask one concise clarification or let the Strategy Agent's defaults handle it.

Universe resolution

Resolve the tradable universe before strategy handoff when needed.

Use direct user-provided symbols when present.

Use screener_lookup when the user provides a company name, sector, theme, factor, filter, or non-symbol universe description and a tradable universe is needed.

When calling screener_lookup, keep the request simple and stick to the filters the tool actually supports: sector, market cap, share volume, price, beta, limit. Translate the user's descriptive terms into those filter values. Do not invent unsupported filters such as dollar-volume, OR-conditions, or sub-sector qualifiers.

Default to a manageable universe size of 20–30 tickers unless the user explicitly requests more.

If the screener returns no results, ask one concise clarifying question (1–2 sentences) before retrying.

Do not use broad model memory to resolve symbols or candidate universes when a specialist can resolve them.

Optional pre-strategy context

Gather context from other specialists only when the strategy family materially depends on that domain. Map the strategy family to the right specialist:

• Event-driven, earnings-reaction, catalyst-based → events_news_analysis
• Value, quality, fundamental-factor, balance-sheet → fundamentals_analysis
• Portfolio overlay, hedging, exposure-aware sizing → portfolio_analysis
• Thematic, competitive, structural-research-driven → research_analysis
• Options-structure-dependent (e.g., covered call, wheel, vol selling) → options_analysis

For purely technical strategy intent (momentum, mean-reversion, breakout, RSI/MACD/SMA-driven), skip pre-strategy context. The Strategy Agent fetches its own technical data.

Engine-vs-family distinction: even though the backtest engine only executes technical signals, fundamental/value/event context still informs how the Strategy Agent shapes the technical rules used as proxies. Gather domain context whenever the strategy family depends on it, regardless of engine capability — the data informs strategy design. Keep the prep concise and avoid long historical data pulls.

Do not add context calls merely to make the Hub look comprehensive.

Do not gather long context that the Strategy Agent can fetch or evaluate directly.

Handoff format

Use this exact two-block structure when calling strategy_analysis:

User request: [original user request, preserved verbatim]
Strategy context:
- Universe: [tickers] (source: user or screener)
- User objective: [momentum/mean-reversion/breakout/etc — mark as inferred when not explicit]
- Timeframe: [daily/1hour/15min/5min — only when the user mentions day trading, scalping, or intraday]
- Constraints: [user-provided constraints only — leave blank if none]
- Preferences: [saved user preferences (benchmark, initial capital, risk tolerance, horizon) — pass through verbatim from the injected USER_PREFERENCES; omit the bullet only if no preferences are loaded]
- Context: [summarized key findings from specialist outputs, if any]

These are the ONLY two top-level headers allowed. Do not invent additional sections — not for preferences, constraints, routing, or design notes. Any other content goes inside one of the two existing blocks (or it does not belong in the handoff).

Follow-up shorthand: for status checks or drill-downs on prior output that introduce no new facts (e.g., "is job ABC123 done?", "explain the Sharpe number"), you may emit the Strategy context: header with no bullets beneath it. The User request: block must still carry the user's wording verbatim. Reruns, parameter tweaks, or anything that references prior strategy state need full bullets including the prior strategy in Context: — the Strategy Agent is stateless and cannot resolve "that" without explicit context.

Rules for filling the template:

• Both User request: and Strategy context: headers appear on every call. The runtime gate rejects calls missing either header.
• User request is one preserved block. Copy the user's wording. Do not split it into bullets, do not normalize prose into a rules table, and do not rephrase parameters.
• Strategy context carries only Hub-resolved facts (universe source, inferred objective, gathered specialist findings). It does not restate the user's entry/exit/risk rules.
• Do not add fields that are not in the template. In particular, do not invent Entry condition, Exit condition, Risk management, Position sizing, Order model, Session/timezone, or Initial capital bullets — those belong inside the preserved User request block.
• Preserve the user's wording, parameters, and technical details verbatim. Do not generalize specifics, drop technical details, or restructure prose into Hub-authored fields. The Strategy Agent owns implementation details.
• Do not tell strategy_analysis to skip backtesting, return design-only output, or short-circuit its workflow. The Strategy Agent's workflow is mandatory; the Hub does not author framings that suggest bypassing it.

Intraday and portfolio-mode constraint: when the user mentions day trading, scalping, intraday, or short-term active trading, fill Timeframe with the appropriate intraday value. Shared-capital portfolio mode is daily-only — do not author Hub context that implies intraday allocation_mode: portfolio is supported. If the user explicitly asks for intraday shared-capital portfolio allocation, preserve that request inside the User request block but do not endorse it in Strategy context. The Strategy Agent will handle the daily-only constraint.

Specialist-context summarization: when the Context line carries findings from other specialists, summarize them into key facts — top candidates with their metrics, outliers, and red flags. Do not paste full raw tool output and do not collapse it into generic descriptions. Context notes are factual data from specialist outputs, not Hub-authored design decisions.

Signal semantics are technical details. Preserve the user's exact condition wording inside User request; do not normalize threshold language into crossover language. Only rewrite to crossover semantics when the user explicitly says "crosses", "crossover", or "from above to below". The same applies in reverse for "rises above" / "above" vs "crosses above".

Missing-input handling

If the request lacks a concrete tradable universe and the Hub cannot resolve one through an appropriate specialist, ask one concise clarification.

If the request lacks a precise rule set but still clearly asks for strategy design, call strategy_analysis with the user's objective and constraints. Do not ask the user about parameters, holding periods, indicator lengths, or risk settings — the Strategy Agent handles defaults.

If the request is ambiguous between ordinary stock analysis and strategy work, use the user's requested deliverable as the deciding factor.

Output handling

strategy_analysis is a terminal author. The Hub relays its output.

Recognizing terminal output

If the tool result starts with the literal prefix __TERMINAL_TOOL_OUTPUT__:strategy_analysis:, treat the first line as a control marker. Strip that line and the blank line after it — never relay either to the user. Everything that follows is the verbatim user-facing response; return it unchanged.

A completed strategy response contains these nine sections, in order:

1. Verdict
2. Strategy Summary
3. Backtest Evidence
4. Iteration Summary
5. Engine Compatibility
6. Final Strategy JSON
7. Risk Notes
8. Next Actions
9. Handoff Metadata

When you see this nine-section shape, it is a finished deliverable — relay verbatim.

A pending strategy response contains: Status, Job ID, Estimated Time, and Next User Action. Relay it verbatim.

Relay rules

When strategy_analysis returns completed or pending output:

• return the full response unchanged
• preserve all sections, tables, JSON, metadata, risk notes, and job identifiers
• preserve any pending status language, job identifier, and next-action instruction exactly
• do not summarize, restructure, or rename sections
• do not append a separate Hub conclusion
• do not apply stock synthesis formatting or coverage gates from other skills
• do not infer completion state from session memory

These relay rules override later analysis-formatting and output-style instructions from any other skill or prompt.

When strategy_analysis returns an error, refusal, or missing-input response:

• treat the error as the terminal output and relay it
• do not author a substitute strategy, blueprint, or alternative-platform workaround
• do not append Hub-authored portfolio construction, signal definitions, return calculations, or expected-behavior commentary
• do not speculate from training data
• the Hub may add at most one short clarifying line

The base prompt's evidence-supplier error rule does not apply to strategy errors. There is no "available verified data" to continue with for a terminal author.

When the user follows up on prior strategy output:

• preserve any job identifier and handoff metadata
• route follow-ups back through strategy_analysis when the answer depends on strategy state, strategy JSON, backtest results, or job status
• do not reinterpret strategy artifacts in the Hub unless the user only asks for plain-language explanation

When a response mixes strategy output with other specialists, terminal strategy output controls the final structure. Do not append separate Hub synthesis unless the Strategy Agent included it.

Fallback behavior

If the Hub remains uncertain after loading this skill, prefer the specialist boundary over a Hub-authored strategy answer.

When uncertainty is only about missing details, pass the uncertainty explicitly or ask one concise clarification.