ai-tools

A collection of AI-agent assets (Claude Code skills, eventually plugins and MCP servers) for working with the Holochain ecosystem.

What's here

skills/holochain-dev

A Claude Code skill that encodes semantic Holochain knowledge — the parts that LLMs get wrong because their training data contains obsolete alpha-version APIs, and the parts that are genuinely non-obvious even when the APIs are current.

What it does (not an exhaustive list):

• Detects when a proposed change affects the DNA hash and, on protected branches, pauses and offers the dev-branch escape hatch rather than silently making the edit
• Verifies HDK/HDI API claims against docs.rs/hdk/<project-version> rather than recall (current target: HDK 0.6.1-rc.5 / HDI 0.7.1-rc.5)
• Inverts the msgpack-version-pin reflex when debugging Rust↔TS serialization errors
• Knows the AgentPubKey-as-EntryHash trick for O(1) joining-timestamp lookups
• Generates sweettest (not tryorama) tests and bootstraps the harness from scratch if the project doesn't have one
• Wraps cargo and test commands with nix develop -c when flake.nix is present
• Detects existing project idioms (coordinator return shape, serialization strategy, signal placement, single-vs-multi-DNA, etc.) and stays consistent rather than imposing a style

Installation: copy skills/holochain-dev/ into ~/.claude/skills/ or wherever your Claude Code setup looks for skills. Eventual plan is to ship as an installable plugin.

notes/

Research and design documents from building the skill:

• holochain-skill-planning.md — original handoff notes
• research-app-patterns.md — pattern survey of unyt-app, dino-adventure, emergence
• research-mewsfeed-friction.md — Claude friction patterns mined from mewsfeed session history
• research-vines-acorn-patterns.md — orthogonal pattern data from vines and acorn
• skill-design-v1.md — synthesis and design brief, with resolved decisions and iteration-2 corrections

These are preserved so future iterations of the skill have the context that shaped it.

Status

Experimental. First deliverable (holochain-dev skill) has been eval-tested through two iterations against five representative Holochain dev prompts, with a measurable pass-rate lift over vanilla Claude Code on the tests where the skill is load-bearing (DNA tripwire, scratch-space semantics, host-fn cost model).

Aspiration

If the skill proves useful in practice, donate to the holochain GitHub org so it can be maintained alongside the HDK.