How to use the Strategy Development Manager agent skill

What is an agent skill, and what does this one make the agent do?

An agent skill is a folder containing a SKILL.md file and supporting files that tell an AI agent how to carry out a particular kind of task. When the skill is installed, the agent can follow its workflow instead of improvising each step.

This skill covers the full path from a paper to a monitored artifact. The agent reads the document, pulls out formulas and rules, writes code that implements them, runs a backtest, checks the results against statistical thresholds, and records each artifact. It does not do the heavy work itself. It delegates to existing tools in the Vibe-Trading system such as read_document, factor_analysis, backtest and alpha_bench.

• Phase 1 INGEST: read and classify a paper as factor-research, strategy or mixed
• Phase 2 EXTRACT: turn the paper into a structured factor or strategy definition and register it
• Phase 3 IMPLEMENT: write the signal engine, run the backtest and link the results
• Phase 4 EVALUATE: compare results against thresholds and mark the artifact active or disabled
• Phase 5 MONITOR: scan active artifacts for decay over time

What do you need before installing?

This skill is not a standalone calculator. Most steps depend on the Vibe-Trading agent environment and its tools. The source lists the following requirements.

• A Vibe-Trading agent environment with the tools read_document, factor_analysis, backtest, alpha_bench, create_hypothesis, sdm_register and sdm_status
• Python with pandas, used by the generated signal engines
• Historical OHLCV market data for the universe you choose
• Optional: an OCR engine for scanned papers, such as rapidocr_onnxruntime, or a vision-capable LLM configured through VIBE_TRADING_OCR_ENGINE

How do you install the skill?

Download the ZIP for this skill from this page and unzip it. Put the resulting strategy-dev-manager folder, which must contain SKILL.md, in one of these locations. Use ~/.claude/skills/ to make it available across your projects, or a project's .claude/skills/ folder to limit it to that project.

If you use the Claude apps instead, open Settings, go to Capabilities, then Skills, and upload the skill there.

• Personal install: ~/.claude/skills/strategy-dev-manager/
• Project install: .claude/skills/strategy-dev-manager/ inside your project
• Claude apps: Settings → Capabilities → Skills, then upload

How do you trigger the skill with natural requests?

You do not need a special command. Describe what you want in plain language and the agent routes the request to the matching phase. The source maps these request types to phases.

If your request covers several phases, such as reading a paper and building a factor, the agent runs them in order from INGEST through EVALUATE.

• Give a paper or report path: starts Phase 1 INGEST
• 'Extract factors from this paper': Phase 2 EXTRACT
• 'Implement and backtest' or 'run the backtest': Phase 3 IMPLEMENT
• 'Evaluate results' or 'check if it works': Phase 4 EVALUATE
• 'Check decay' or 'monitor factors': Phase 5 MONITOR
• 'Disable factor' or 'enable factor': changes the artifact status
• 'List my factors' or 'show status': lists tracked artifacts

What does a worked example look like?

The bundled examples.md uses Jegadeesh and Titman (1993), 'Returns to Buying Winners and Selling Losers'. The agent classifies the paper as factor-research and extracts a momentum factor named momentum_12_1, with a 12-period formation window and a 1-period skip. It records the target universe as US equities.

The extracted definition is then registered, implemented in a signal engine, backtested and evaluated. The source does not give a final backtest result for this example, so do not expect one from the skill's documentation alone.

What thresholds does the skill use to judge results?

For factors, the skill checks IC mean above 0.03, IR above 0.5 and an IC positive ratio above 55%. For strategies, it checks a Sharpe ratio above 0.5, a max drawdown below 30%, and reports win rate and profit factor for context.

Before implementing, the agent also checks for duplicates. A Pearson IC above 0.99 against an existing alpha is treated as a duplicate and stopped. An IC between 0.90 and 0.99 may be kept as a variant with a note. These are simple rules of thumb set in the skill, not proof that a factor will work.

• Factor alive: IC mean > 0.03, IR > 0.5, IC positive ratio > 55%
• Strategy alive: Sharpe > 0.5, max drawdown < 30%
• Dead artifacts are set to disabled with a recorded reason

How does decay monitoring work?

The skill runs sdm_decay_scan across active artifacts in a universe and labels each one healthy, warning, decayed or critical. Each metric is judged on its own, and the worst signal sets the overall label. Status moves to monitoring after three consecutive periodic warning-or-worse results, and to decayed after a decayed-or-worse run.

The source describes scheduled decay scans run by a background executor. That executor is off by default and must be turned on with an environment flag, which the source documents in references/scheduled_decay_scan.md.

• IC ratio (rolling/baseline): healthy above 0.7, critical below 0.3
• IR (rolling window): healthy above 1.0, critical below 0.1
• Sharpe (strategies only): healthy above 1.0, critical below 0.0

What safety points should you know?

The skill runs scripts and uses the network. It writes and executes strategy code, which may contain errors or lookahead bias. Review the generated signal_engine.py before trusting any backtest. Backtest results on historical data do not predict future performance, and the skill does not provide live trading advice or guaranteed returns.

The skill's own validation step only checks that the file parses as Python. It does not check the logic.

• Read the generated code before running or acting on it
• Check the backtest for lookahead bias and data errors
• Do not connect the output to automatic real-money trading without review