Skip to content

Agent skills

A general-purpose language model does not reliably know the details of the Intuition protocol. The Intuition agent skills supply those details as verified reference material that an agent loads when a task needs it. They follow the Agent Skills specification (opens in a new tab) and work with Claude Code, Codex, and other compatible agents.

What the skills prevent

Agents that work without the skills tend to make the same mistakes:

  • They cannot discover the contract ABIs, because general block explorers do not index the Intuition network.
  • They treat term identifiers as integers. Identifiers are bytes32 values.
  • They try to create a single term. Creation is batch-only: createAtoms and createTriples take arrays.
  • They omit or miscalculate the value a transaction must carry.

Available skills

SkillPurpose
intuitionThe reference for protocol transactions: ABIs, encoding, addresses, and value calculations
erc8004-agent-layerPlans an ERC-8004 trust-provider integration. See ERC-8004 agent layer

The skills are published in the 0xintuition/agent-skills (opens in a new tab) repository.

Install

You need Node.js and a package manager. The examples use pnpm.

  1. Install every skill:

    pnpm dlx skills add 0xintuition/agent-skills

    To install one skill, name it:

    pnpm dlx skills add 0xintuition/agent-skills --skill intuition

  2. Start a new agent session. The installed skills are available in it.

Invoke a skill by name, for example /intuition. Agents that support skills also invoke one on their own when a task matches it.

What the intuition skill produces

The intuition skill reads current costs from the contract, builds the calldata and value, and simulates the transaction. It then outputs one unsigned transaction as JSON:

{
  "to": "0x<multivault-address>",
  "data": "0x<calldata>",
  "value": "<amount in wei, as a base-10 string>",
  "chainId": "<chain ID, as a base-10 string>"
}

The skill does not hold keys and does not send transactions. Your application, wallet, or backend remains responsible for:

  • choosing the signer,
  • presenting the transaction for approval,
  • broadcasting it, and
  • handling the receipt.

This boundary is deliberate. An agent can prepare a correct transaction without ever being in a position to authorize one.

Approval policy for unattended agents

An agent that runs unattended reads its approval policy from ./.intuition/autonomous-policy.json, or from the path in INTUITION_POLICY_PATH. Without a policy file, the skill uses manual-review mode. When the policy requires review, the skill outputs an approval request with "status": "approval_required" instead of an executable transaction. Enforce the policy in your signer as well. The skill does not include signing middleware.

Pin a version

The repository's main branch is an integration branch, not a stable channel. For production agent workflows, install from a release tag or a commit, and upgrade deliberately. Release tags are named after the skill and its version, in the form <skill>-v<version>. The repository's CHANGELOG.md lists the changes in each release.