emlai/defi-native-skill

What it claims

defi-native is an MIT-licensed Agent Skill. It gives a coding agent two things. The first is a set of evergreen capital-markets mental models. The second is a discipline of pulling live data before stating any number, with an as-of date on every figure. Its example output for a vault breakdown:

• names five layers
• looks through to the real collateral
• names the oracle class
• splits base yield from incentives
• identifies first loss
• says whether liquidations can fire on the price humans see

The author says the corpus behind it is over 6,000 pages. It includes banking and finance books, 2026 conference transcripts (TokenizeThis NYC, Vault Summit, Stablecoin Summit, a CFTC advisory committee meeting), podcasts, SEC filings and papers. The diagram cites 140 verified doc sources, while the prose says "90+" and "130+". Every URL in the manifest was liveness-checked.

How it works

Only the skill description is always loaded. SKILL.md holds 9 prime directives and a routing step. It loads when a DeFi question triggers the skill, classifies the request, and loads one reference file:

flowchart TD
    S[SKILL.md classifies ask] -->|learn| A[analogs.md / task-playbooks.md]
    S -->|assess / recommend| P[defi-opportunities-playbook.md]
    S -->|options / LP| O[options-and-liquidity.md]
    S -->|rate a manager| C[curation-frameworks.md]
    S -->|monitor| M[market-pulse.md]
    P --> K[checklist.md: unanswered = finding]
    S --> R[api-routes.json]
    R -->|MCP, then keyed, then keyless| D[live pulls / scripts/pulse.py]

Other references cover trades, RWA fund mechanics, tokens, perps, credit cycles, an 18-section concepts.md and a 136-entry glossary.

• Assessment playbook: 12 steps.
• Curator rating: eleven frameworks folded into one scoreable anatomy.
• Data routing: api-routes.json prefers MCP, then keyed APIs, then keyless ones.
• Manifest use: the agent fetches only 4–6 rows from manifest.json, never the whole file.
• Recommendations: each must include decomposed risk, labeled-basis probabilities, risk:reward including the total-loss branch, invalidation triggers and a runner-up.

Evidence

Before release, the skill went through three checks:

• paired runs with and without the skill, graded on structure rather than memorized numbers
• an adversarial audit by a panel of independent models
• re-verification of load-bearing claims against primary sources

No scores, win rates or eval sizes are published. Two worked outputs are provided:

• examples/assessment-example.md: a fictional product showing the minimum-bar skeleton
• evals/sample-output-2026-08-28.md: a real run under the recommendation protocol

How to use it

Install with:

npx skills add emlai/defi-native-skill

This needs Node.js. The CLI detects the host agent. The skill triggers automatically on DeFi questions in a new session. To install manually:

git clone https://github.com/emlai/defi-native-skill.git
ln -s "$(pwd)/defi-native-skill" ~/.claude/skills/defi-native

For other hosts, put the folder in that host's skills directory. The cross-agent standard location is .agents/skills/. Update with npx skills update or git pull.

All core features work without API keys. Optional keys are listed in references/data-sources.md.

Contributions must follow strict conventions in CONTRIBUTING.md: dated numbers, no em dashes, and no recommendations without the full view.

Limits

• No eval numbers: the eval gate is described but no results are published, so its effect cannot be checked from the page.
• Dated figures: numbers inside the skill are calibration examples from 2026 and must be re-verified at use time.
• No auto-update: the skill only reports newer versions during monitoring tasks. It never loads remote files at runtime, so what you reviewed is what runs.
• Read-only: it never signs, submits or approves transactions.
• Not advice: output is labeled research, not financial advice, and every assessment states that DeFi carries total-loss tails.