Cardano Dev Skills Connects AI Agents to Current Documentation

Cardano Dev Skills gives AI coding agents instructions and technical references for building transactions, writing smart contracts, and understanding Cardano standards. The package addresses a recurring development problem: agents can produce convincing code that relies on outdated libraries or network rules.

Giovanni Gargiulo, Senior Blockchain Architect at the Cardano Foundation, demonstrated the tool during Developers Office Hours on September 18, taking an agent from a CIP-68 explanation to a Java implementation example.

Aiken and Conway Errors in Cardano Code

Gargiulo began the project after finding that correcting AI-generated code sometimes took longer than implementing the solution himself. What started as a personal experiment later moved into the Cardano Foundation’s GitHub organization and opened to community contributions.

One recurring issue involved Aiken, a language for developing smart contracts on Cardano. The agent continued using patterns from standard library version 2.4 even when Gargiulo requested documentation for version 3. The resulting code failed to compile.

Conway-era changes created a different failure point. When working with stake address registration and certificates, the agent combined newer requirements with an older transaction construction approach. The application code could compile successfully while the transaction failed when submitted to the network.

Software development kits, or SDKs, presented similar problems. These libraries allow applications to construct transactions and interact with Cardano, but renamed fields or changed function signatures can make older examples unusable.

Gargiulo initially downloaded current documentation manually and repeatedly directed the agent to it. Sessions restarting or conversation context being compressed meant those instructions had to be supplied again. His experiments with documentation retrieval also required ongoing source maintenance, leading to a package that combined references with reusable development workflows.

Maintained Sources and the CIP-68 Demonstration

The version presented during the session contained 15 development skills and 55 documentation sources. On October 2, the project’s official website listed 18 skills and 73 sources, covering SDKs, programming languages, infrastructure, CIPs, and ledger specifications.

A skill provides instructions for a specific task, such as building an ADA transfer transaction, writing a validator, or optimizing a contract. Cardano Improvement Proposals, known as CIPs, supply technical standards that agents can explain and apply through the accompanying sources.

The package separates its source registry, task instructions, and downloaded references. Maintainers can update documentation independently of the workflows that tell an agent how to use it.

Task descriptions also help the agent match ordinary requests to the appropriate workflow. A developer can ask for a token transaction or contract review in natural language, allowing the agent to select the relevant skill without an explicit command.

An automated weekly process retrieves upstream changes and opens a GitHub pull request for human review before they enter the main branch. Users then obtain the updated package. Claude Code users must enable automatic updates for its marketplace or update it manually.

The cardano-context skill adds a persistent project instruction telling Claude Code or Codex to treat existing Cardano knowledge as potentially outdated and consult the bundled references first. Both agents use the same development skills and sources through their respective installation routes.

In the demonstration, Gargiulo requested an explanation of CIP-68, the metadata standard for Cardano native assets. It distinguishes a reference NFT, whose transaction output holds the metadata, from a user token representing the asset in a holder’s wallet. The issuer’s contract logic determines whether the metadata can be updated.

The agent then responded to a request for a corresponding transaction by selecting development instructions and presenting available libraries. Gargiulo chose cardano-client-lib, a Java library, and received project configuration, dependencies, and implementation code. The generated example specified Java 17.

Gargiulo did not execute the example during the session. The demonstration therefore covered preparation of the implementation, with end-to-end transaction testing still outstanding.

Feedback and Further Cardano Applications

Cardano Dev Skills includes a feedback mechanism for reporting incorrect documentation, missing instructions, or failures to select an appropriate skill. Bora, a project contributor, described an instance in which an agent proposed a report at the end of a session and asked the user to review it before approving submission as a GitHub issue.

Maintainers use those reports to assess whether a source or workflow needs correction. Gargiulo also distinguished documentation problems from requests that are themselves unclear or technically flawed, since each requires a different response.

Specialist contributions help expand coverage. Gargiulo cited the addition of Go material by a contributor familiar with the language and its Cardano tooling. The project’s scope rules admit development integrations, specifications, SDKs, and infrastructure documentation while excluding marketing material and end-product user manuals.

The package’s text can be inspected directly, while its helper hooks require review because they execute code. Gargiulo outlined further sessions on prompt injection, trust boundaries, and situations in which agents ignore documentation. He also described plans for additional work on modularity and compatibility with other agents in the second half of October.

Potential extensions discussed during the session included assessing governance proposals against the Cardano Constitution and a testnet experiment in which users would pay an agent in USDCx for a smart contract review, potentially through Masumi. These were presented as work in progress and a proposed experiment, respectively.

Hackathon onboarding supplied a concrete application for the existing package. For a planned 36-hour event in Singapore, a participant described giving developers Cardano Dev Skills and a few starting prompts rather than assembling an extensive resource list for every team.

In that workflow, a team can begin with a transaction, validator, or token standard inside its coding agent. The relevant task instructions and technical references are available in the same environment, while problems discovered during implementation can be reported through the package’s feedback process.