Skills - Helius Docs

Skills are model-agnostic instruction sets that teach AI assistants how to route your requests to the right MCP tools and reference files. They go beyond raw tool access — they include routing logic, correct SDK patterns, and rules that prevent common mistakes like hardcoded fees, wrong endpoints, or missing Sender tips. Skills work with any AI platform — Claude Code, Codex, OpenAI API, Claude API, Cursor, ChatGPT, and more. Each skill ships with pre-built system prompt variants so you can drop them into whatever tool you use.

Available Skills

Build

Core Solana development — DAS, Sender, WebSockets, LaserStream, Webhooks, Wallet API, and more

Phantom

Frontend dApp development — Phantom Connect SDK (React, React Native, browser), token gating, payments

DFlow

Trading applications — spot swaps, prediction markets, real-time streaming, Proof KYC

SVM

Solana protocol internals — SVM execution, consensus, account model, validator economics

How Skills Work

Skills follow a 3-layer prompt architecture:

Layer	What	Where
A: Harness	Agent/runtime-specific behavior and tool rules	Built-in (Claude Code), `AGENTS.md` (Codex), preamble in prompt variants (API)
B: Skills	Canonical, reusable, model-agnostic procedures	`SKILL.md` files + `references/` directory
C: Task	User request + dynamic context	Provided at runtime by each platform

The skill content (Layer B) is identical across platforms — only the harness wrapper (Layer A) differs. Each skill includes:

Routing logic — maps your request to the correct MCP tools and reference files
Reference files — deep documentation for each Helius product the AI reads before writing code
Rules — prevents common mistakes (e.g., hardcoded priority fees, wrong RPC endpoints, missing error handling)

Using Skills Across Platforms

Each skill ships with three pre-built prompt variants in its prompts/ directory:

File	Use Case	Format
`openai.developer.md`	OpenAI Responses / Chat Completions API	Layer A preamble + skill content with `=== BEGIN SKILL ===` delimiters
`claude.system.md`	Claude API system prompt	Layer A preamble + skill content with delimiters
`full.md`	Cursor Rules, ChatGPT custom instructions, other tools	All reference files inlined, no external dependencies

Find these in:

Repo: .agents/skills/<skill>/prompts/
npm package: helius-mcp/system-prompts/<skill>/

Install skills directly into your AI coding agent using the Vercel Skills CLI. Works with Claude Code, Cursor, Copilot, Codex, and more.

# Install a single skill
npx skills add helius-labs/core-ai --skill build

# Browse all available Helius skills
npx skills add helius-labs/core-ai --list

Skill	Install Command
build — Core Solana dev (Sender, DAS, WebSockets, LaserStream, webhooks)	`npx skills add helius-labs/core-ai --skill build`
dflow — Trading apps (swaps, prediction markets, streaming, KYC)	`npx skills add helius-labs/core-ai --skill dflow`
phantom — Frontend dApps (Phantom Connect, React, token gating)	`npx skills add helius-labs/core-ai --skill phantom`
svm — Protocol internals (SVM engine, consensus, account model)	`npx skills add helius-labs/core-ai --skill svm`

The Helius plugin bundles all skills + MCP servers in one install:

/plugin marketplace add helius-labs/core-ai
/plugin install helius@helius-labs

Skills are invoked with /helius:build, /helius:phantom, /helius:dflow, or /helius:svm.

Each skill can be installed independently using its install.sh script:

git clone https://github.com/helius-labs/core-ai.git
cd core-ai/helius-skills
./helius/install.sh            # Install to ~/.claude/skills/helius/
./helius/install.sh --project  # Or install to current project

Standalone skills are auto-invoked as Agent Skills — no slash command needed.

Codex auto-discovers skills from .agents/skills/ in your project root:

# Copy the .agents/ directory into your project
cp -r /path/to/core-ai/.agents/ your-project/

Skills are triggered implicitly when your task matches, or explicitly with $helius, $helius-dflow, $helius-phantom, or $svm.

Use openai.developer.md as a developer message:

from openai import OpenAI

client = OpenAI()

with open(".agents/skills/helius/prompts/openai.developer.md") as f:
    skill_prompt = f.read()

response = client.responses.create(
    model="o3",
    instructions=skill_prompt,
    input="Build a swap interface that submits via Helius Sender"
)

Also works with Chat Completions using {"role": "developer", "content": skill_prompt}.

Use claude.system.md as the system prompt:

import anthropic

client = anthropic.Anthropic()

with open(".agents/skills/helius/prompts/claude.system.md") as f:
    skill_prompt = f.read()

message = client.messages.create(
    model="claude-sonnet-4-6-20250514",
    max_tokens=4096,
    system=skill_prompt,
    messages=[
        {"role": "user", "content": "Build a swap interface that submits via Helius Sender"}
    ]
)

Use full.md — it inlines all reference files so there are no external dependencies.Cursor: Copy to .cursor/rules/helius.md:

cp .agents/skills/helius/prompts/full.md .cursor/rules/helius.md

ChatGPT Custom Instructions: Use the core SKILL.md content (the full.md may exceed ChatGPT’s 8,000 character limit for larger skills).

Prerequisites

All skills require the Helius MCP server to be running (or the equivalent tools configured in your agent framework). The DFlow skill additionally requires the DFlow MCP server. The Phantom skill additionally requires a Phantom Portal account for embedded wallets.

Skill	Requires	API Key Needed
Build	Helius MCP	Yes (Helius)
Phantom	Helius MCP	Yes (Helius)
DFlow	Helius MCP + DFlow MCP	Yes (Helius + DFlow for production)
SVM	Helius MCP	No

For Agents

​Available Skills