Investment Agent

Investment Agent Skill

This skill supports two deployment styles. Pick one based on the user's intent.

Command Invocation

Treat /investment-skill as the primary explicit invocation alias. The text after the command is the user request. Also support $investment-skill, /investment-agent, and $investment-agent as aliases.

Default to Mode B for command-style invocations unless the user explicitly asks to run a Telegram bot, background service, scheduler, server, or deployment.

Command routing:

• /investment-skill 금일 리포트 작성, /investment-skill 오늘 리포트, /investment-skill daily report: run scripts/build_agent_request.py report.
• /investment-skill 오늘 반도체 섹터 어때?, stock/sector/macro questions, or general investment questions: run scripts/build_agent_request.py chat "".
• /investment-skill 포지션 점검 ...: if the user provides complete position JSON or enough fields to build it, run scripts/build_agent_request.py position --position-json ''; otherwise ask for the missing ticker, market, entry price, quantity, direction, and current price if needed.

When using Mode B, read the returned JSON payload and use its system_prompt, user_prompt, and data as context for the final answer. The payload is an internal handoff format, not the user-facing result. Do not expose the raw payload unless the user asks for it. Report prompts contain a REPORT_INPUT_JSON object with a deterministic execution layer; use its action_status, risk_gate_status, is_executable, and price/risk fields as authoritative.

For report requests, produce a human-readable Korean Markdown report. Do not answer with raw JSON. Include a one-line conclusion, market overview, sector/theme summary, an action plan based only on deterministic execution statuses, wait/do-not-chase candidates, portfolio strategy, key risks, data-quality notes, and an investment disclaimer. Do not invent prices, targets, stops, position sizes, or execution permissions that are absent from price_engine_output, risk_gate_results, or the returned report data.

Mode A: Standalone Telegram System

Use this when the user wants a 24/7 bot server.

• Requires TELEGRAM_BOT_TOKEN, TELEGRAM_CHAT_ID, and an OpenAI-compatible OPENAI_API_KEY.
• Runs python run.py or the systemd service template.
• Telegram is the UI. The bot collects data, calls the configured AI API, stores reports/positions locally, and sends replies or alerts.
• Entry point: src/bot.py.
• AI API entry point: src/ai_client.py.

Setup:

cp templates/config/.env.example .env
pip install -r requirements.txt
python run.py

Mode B: On-Demand Agent Command

Use this when Codex, OpenAI Agents SDK, Manus, or another agent runtime should act as the AI.

• Does not require Telegram credentials.
• Does not start a long-running scheduler or background automation.
• Builds market context, enriched news context, data-quality checks, and deterministic execution prompts only when the user asks.
• The outer agent should read the returned system_prompt and user_prompt, then answer with its own model.
• Entry point: src/agent_adapter.py.
• CLI helper: scripts/build_agent_request.py.

Examples:

scripts/build_agent_request.py report
scripts/build_agent_request.py chat "오늘 반도체 섹터 어때?"
scripts/build_agent_request.py position --position-json '{"id":1,"name":"NVIDIA","ticker":"NVDA","market":"US","entry_price":100,"quantity":2,"currency":"USD","direction":"long"}' --current-price 112

Python:

from src.agent_adapter import build_report_request

request = build_report_request()
# Send request["system_prompt"] and request["user_prompt"] to the outer agent model.

Core Analysis Modules

• src/report_engine.py: collects on-demand market data and builds REPORT_INPUT_JSON prompts.
• src/price_engine.py: calculates entry, stop, target, risk/reward, and position size.
• src/risk_gate.py: blocks unsafe or unsupported execution candidates.
• src/recommendation_safety.py: applies deterministic safety controls to LLM candidates.
• src/data_quality_engine.py: scores stale, missing, or conflicting market data.
• src/news_classifier.py / src/news_impact_engine.py: classify report news and score market impact.
• src/momentum_inflection.py: detects momentum phase and inflection quality.
• src/market_regime_engine.py: classifies Korean and US market regimes and risk budgets.
• src/entry_filter.py: checks RSI, Bollinger Bands, volume, and ADX for entry timing.
• src/position_tracker.py: parses and evaluates positions.

This skill can use local market memory and snapshot/news outcome stores when present, but Mode B remains on-demand: it does not start Telegram, a scheduler, or a background service unless the user explicitly asks for standalone bot mode.

Read references/module-guide.md for module details and references/prompt-engineering.md for prompt rules.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jungducknam
• Source: jungducknam/investment-agent-skill
• License: MIT

Install and usage instructions live in the source repository linked above.