How to use the Strategy Discovery agent skill (strategy-discovery)

What an agent skill is and what this one does

An agent skill is a folder containing a SKILL.md file that tells an AI agent when and how to perform a task. The agent reads the instructions when a request matches them.

Strategy Discovery is a single entry point over two sources: the Alpha Zoo registry and the SDM strategy store. Its answers are computed from backtest results. Every returned row shows how fresh its evidence is, so you can tell whether a result is current before relying on it.

• Lists strategies in the catalogue, optionally filtered by source (alpha_zoo or sdm).
• Queries strategies by regime, minimum Sharpe ratio and evidence quality.
• Shows the per-regime evidence rows behind one strategy.
• Rebuilds the local evidence cache from existing backtest run folders.

What the skill cannot do

The three query tools are read-only. The skill cannot create, register, mutate or delete strategies. Requests to create, backtest or register a strategy are routed to other workflows (strategy-generate, strategy-dev-manager or alpha-zoo).

The only write it performs is refreshing the disposable evidence cache from local backtest run artifacts.

Install the skill (ZIP or Claude apps)

Download the ZIP for this skill from this page, then choose one of these routes.

For a personal install, unzip it into ~/.claude/skills/. For a project-only install, unzip it into the project's .claude/skills/ folder. You can also upload the skill in the Claude apps under Settings → Capabilities → Skills.

• Personal: unzip into ~/.claude/skills/
• Project: unzip into .claude/skills/ inside your project
• Claude apps: Settings → Capabilities → Skills, then upload

Know what the skill needs before you start

The skill needs the Vibe-Trading repository with the Alpha Zoo and SDM strategy store available locally. To populate evidence, you also need local backtest run artifacts (run directories). Without run artifacts, the evidence tables have nothing to report.

The skill runs scripts. It does not use the network.

• Vibe-Trading repository with Alpha Zoo and SDM strategy store available locally
• Local backtest run directories containing artifacts/trades.csv and artifacts/equity.csv for evidence
• Runs scripts: yes
• Uses network: no

Trigger the skill with natural requests

You do not need to name tools. Ask in plain language and the agent routes the request. The skill's decision rules map requests like these to the right tool.

• "List available strategies" → list_strategies
• "What works in bear markets?" or "anything with Sharpe above 1?" → query_strategies
• "Show the evidence behind strategy X" → get_strategy_evidence
• "The evidence is stale, refresh it" or "turn my backtest runs into evidence" → refresh_strategy_evidence
• "Create or backtest a new strategy" → not this skill; routed to strategy-generate, strategy-dev-manager or alpha-zoo

Refresh the evidence cache from backtest runs

Refresh reads run specs from a JSON manifest, either an object with a runs array or a bare array of run specs. Each spec needs a non-empty strategy_id and a run_dir, with an optional position_size. You supply exactly one of manifest_path or runs.

Entries that fail the hard gates or fall outside allowed paths are skipped with a stable reason, and the rest still process. The result reports status, runs, strategies, rows and skipped.

The equivalent CLI command is shown below.

• Exactly one of manifest_path or runs is required
• Each spec: strategy_id, run_dir, optional position_size
• Skipped reasons include hard-gate:* and path-outside-allowed-roots:

Read the evidence fields

Every row carries its own metadata. evidence_stage tells you how far the evidence goes; the harness currently writes only backtest, so a row means backtest evidence exists, not holdout, shadow or live evidence. provenance points to the run directory the row came from.

decay_status (fresh, aging or stale) is computed at read time from evidence_age_days. staleness_days is reported only and never blocks a result. breakeven_fee_bps can be null when the cost screen cannot be verified.

• evidence_stage: backtest is the only stage written today
• provenance: the source run directory
• decay_status: fresh, aging or stale
• evidence_age_days drives decay_status
• breakeven_fee_bps is null when unverifiable

Use query filters carefully

By default, query_strategies applies a minimum of 10 trades, keeps only cost-feasible rows, requires adequate evidence quality and excludes stale rows. Each default is a gate you can relax deliberately.

Setting min_evidence_quality to any only removes the quality floor; rows must still pass the trade, cost and Sharpe filters. Setting include_stale to true surfaces stale rows with their stale-evidence warnings, sorted after fresh ones.

• Defaults: min_trades 10, cost_feasible true, min_evidence_quality adequate, include_stale false, limit 10
• cost_feasible false shows rows with unverifiable breakeven and their warnings
• get_strategy_evidence returns all rows for a strategy without filtering