Open-source library of 501 tested AI agent skills for trading infrastructure hits Hacker News

What the library contains

An open-source project called Algo-Trading-Skills reached the Hacker News front page as a Show HN post dated 11 September 2026. Built under the GitHub handle HimanshuJ16, the repository packages 501 skills for AI coding agents — structured instruction documents following the agentskills.io open standard, each paired with a working Python reference implementation and its own unit tests. According to the project's page, the full library is backed by 20,291 tests that run in continuous integration.

The skills span 16 engineering domains, covering broker API integration, backtesting, risk management, execution algorithms, machine learning and compliance, among others. Individual skills are named for concrete failure modes rather than abstract topics: idempotent order placement to avoid duplicate executions after a reconnection, drawdown-triggered kill switches, self-detection of wash trades and spoofing, and velocity limits that freeze a hot wallet when withdrawal anomalies appear. The project's core argument is that a language model can produce trading code that looks correct — clean structure, plausible logic — yet still fails in production for reasons no syntax check catches, such as a backtest quietly using a bar's own closing price to predict that same bar's direction, or a callback blocking the WebSocket read loop during the volatility spike a strategy was built to capture.

Regulatory mapping

According to the repository, the library maps onto five regulatory and exchange framework groups: US SEC and FINRA rules (including Rule 15c3-5 pre-trade risk controls, Reg NMS order protection and Reg SHO short-sale locates), the EU's MiFID II, RTS 6 and MAR market-abuse regime, the UK FCA's algorithmic trading controls and Senior Managers regime, global regulators such as ASIC, SEBI, MAS and IIROC, and ISDA standards for OTC derivatives. Example skills cite broker and platform touchpoints including Alpaca, IBKR, Zerodha, Fyers, Upstox, Coinbase, Fireblocks and BitGo.

Every skill also carries a "When NOT to Use" section that names the cases it does not cover and points to the skill that handles them. The project frames this as a guard against a specific failure mode: an agent correctly applying a playbook to the wrong problem, which keyword matching is particularly prone to causing.

Compatibility and context costs

The skills work with Claude Code, GitHub Copilot, OpenAI Codex CLI, Cursor and Gemini CLI — in practice, anything that reads SKILL.md files. Installation options include a per-domain plugin marketplace for Claude Code, an interactive npx skills add command, and a plain git clone with validation scripts that check structure, frontmatter and cross-references before running every skill's test suite.

The packaging decision is notable. The project ships one plugin per engineering domain rather than a monolithic bundle, because Claude Code loads the name and description of every installed skill into the model's context at the start of each session. A single domain costs a few thousand tokens, while the all-in-one plugin runs to tens of thousands per session. One disk-level caveat remains: installing one domain still checks out the full repository, roughly 100 MB, though only that domain's skills load into context.

Explicit limits and disclaimers

The project describes itself as an independent community effort with no affiliation to Anthropic PBC or any broker, exchange or vendor named in the repository. It states plainly that the skills encode production engineering practices, not financial, legal, tax or compliance advice, that they do not guarantee strategy profitability and that they do not remove the risk of capital loss in live trading. The recommended path is paper accounts or live environments with strictly enforced risk limits. The code is Apache 2.0 licensed.

Why it matters

The repository is a clear example of an emerging pattern: encoding specialist practitioner knowledge as versioned, machine-readable skill files that agents load on demand, instead of hoping a general-purpose model has absorbed the operational details of a domain. The verification angle is what distinguishes it — each skill ships tests rather than prose alone, and a roughly 20,000-test CI suite gives a concrete, if self-reported, quality signal. It is also a useful data point on context economics: skill libraries are now being designed around per-session token costs, with domain-level packaging as the default rather than a single mega-bundle. For teams building trading or other regulated infrastructure with AI agents, the structure here — scoped skills, explicit non-applicability sections, regulatory cross-references — may be as instructive as the trading content itself.