Every prompt repo on GitHub follows the same pattern. "Act as a senior software engineer. Be helpful and thorough." That's not a skill โ€” that's a costume you slap on Claude before sending it into battle. A developer posting on DEV.to has spent six months building something fundamentally different: AgentOS 2.0, a collection of 135 production-grade Claude Skills that embed actual domain formulas, named sub-agents with distinct responsibilities, and explicit forbidden behaviors designed to prevent hallucination in critical areas.

The Three Traps That Kill Every Prompt Repo

The author identifies three patterns that make most prompt collections useless in practice. The "costume prompt" assigns a generic role like 'expert financial analyst' without any framework underneath. The "instruction dump" piles on behavioral requests โ€” be professional, cite sources, consider multiple angles โ€” which just describes what Claude already does. And the "persona prompt" invents a named character with fake credentials, relying on roleplay instead of actual expertise. Real production-grade skills need sub-agents with named responsibilities, domain formulas written in executable code, and explicit constraints that stop the model from taking the easy wrong path.

FinanceOracle: The Apex Skill That Actually Runs

FinanceOracle is the most complete skill in the collection, built around 12 distinct sub-agents including OptionsDesk for derivatives pricing, MacroStrategist for macro regime analysis, HedgeFundArchitect for strategy design, and FamilyOfficeCIO for multi-generational allocation. Rather than generic advice, it ships real Python: a Black-Scholes implementation that returns price, delta, gamma, vega, and theta by calculating d1 and d2 values from the strike price, time to expiration, risk-free rate, and implied volatility. It also includes a Black-Litterman portfolio construction function that computes equilibrium returns using market capitalization weights, view matrices, and uncertainty parameters โ€” code you can paste into a notebook and run immediately.

OKREngine Catches Failure Modes Before They Kill Your Quarter

The OKREngine skill exists because the author watched two startups waste entire quarters on broken objectives and key results. Its objective quality scorer penalizes vague language like 'improve' at the start of an objective, flags aspirational fluff such as 'best in class,' and rewards brevity under 100 characters. For each key result, it checks for presence of a metric, baseline value, and target โ€” awarding points when all three exist. The skill also catches the 12 most common OKR failure modes including sandbagging, health metrics disguised as business results, and the single most destructive mistake: tying OKR scores to compensation.

VentureIntelligence Turns Term Sheets Into Structured Red Flags

The VentureIntelligence skill includes a term sheet scoring function that flags specific deal terms with severity ratings. It catches liquidation preference multiples above 1x, participating preferred structures where VCs get paid twice in exits below threshold, full ratchet anti-dilution provisions that become catastrophic during down rounds, and board composition imbalances where investor seats outnumber founder seats โ€” the last one meaning founders can be fired from their own company. The function returns a score from 0 to 10 with grades ranging from 'Sign it' down to 'Get a lawyer NOW.'

CrisisIntelligence Provides a War Room Operating System

Every company will face a crisis; almost none prepare for one. The CrisisIntelligence skill includes a classification function that scores severity across customer impact percentage, revenue at risk in dollars, media coverage level, regulatory involvement flags, and legal liability indicators. A score above 70 triggers CRITICAL (P0) status with CEO leadership and immediate war room activation. Below 40 falls to MEDIUM (P2) handled at the director level with external monitoring. The skill embeds a 5Rs framework โ€” Recognize, Respond, Responsible, Remediate, Restore โ€” baked into every communication template generated.

Setup Takes About 60 Seconds

Claude.ai Projects users paste the SKILL.md file directly into Project Instructions and start chatting. Claude Code users append the skill file to their CLAUDE.md with a single shell command: cat finance-oracle/SKILL.md >> .claude/CLAUDE.md. For API integration, developers read the skill markdown into a string variable and pass it as the system prompt alongside any model from the claude-sonnet-4 series upward. The entire catalog spans 135 skills across ten categories including startup management (11), AI and engineering (14), data and analytics (10), finance (9), operations and business (20), product and customer (11), developer tools (19), and specialized domains like healthcare, biotech, cybersecurity, and real estate (12).

Key Takeaways

  • Most Claude prompt repos are costume prompts โ€” they describe a role without providing the frameworks that make experts actually useful
  • AgentOS 2.0 embeds runnable Python code including Black-Scholes derivatives pricing and Black-Litterman portfolio construction directly in skill definitions
  • The OKREngine scores objectives on structural quality markers and catches 12 documented failure modes before teams commit to broken goals
  • VentureIntelligence flags specific term sheet provisions โ€” like full ratchet anti-dilution or investor-majority board composition โ€” with actionable severity ratings
  • Skills are MIT licensed, free forever, and work across Claude.ai Projects, Claude Code, and the API in under a minute of setup

The Bottom Line

If you've been burning time prompting Claude to 'be more like a senior developer' or 'act like a CFO,' you're doing it wrong. This repo proves that AI skills need actual infrastructure โ€” formulas, sub-agents, forbidden behaviors โ€” not pep talks dressed up as system prompts. For builders who want tooling that actually works on day one rather than another collection of vibes-based roleplay templates, AgentOS 2.0 is worth the hour it takes to explore.