anzal1

Chariot — AI skill for Claude Code

AI community

A harness of harnesses: every AI agent you use pulls one load, with you holding the reins.

How to install Chariot

This entry records only its repository, not the path inside it, so there is no exact command to give. Open anzal1/chariot and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Chariot does

A harness of harnesses: every AI agent you use pulls one load, with you holding the reins.

Alternatives in AI

Impeccable — The design language that makes your AI harness better at design 13.9k ★
ARIS Agent Guide — For AI agents reading this repo 6.2k ★
Osaurus — Own your AI. The native macOS harness for AI agents -- any model, persistent memory, autonomous execution, cry 5.1k ★

README

Chariot

Every AI agent you use is a harnessed horse. Chariot is what makes them pull one load, with you holding the reins.

A harness of harnesses: one local layer that any harness, SDK, model CLI, API or website can join. It hands out tasks, keeps members from wiping each other's work, fails over when a subscription runs out, and checks what was really done. It has no model of its own and never holds anyone's keys.

Status: design, tested, not built yet.

File	What it is
ARCHITECTURE.md	The engine: core, protocol (Chariot/1), drivers, onboarding, every test result
ABSTRACTIONS.md	The user and developer layer on top: member, task, hold, rule; `chariot.yaml`; extension points
diagrams/architecture.html	The architecture as diagrams
experiments/	The rigs behind every number in the docs

What the experiments showed (2026-10-12)

Test	Result
Rules taught in the prompt channel vs only in tool descriptions	18/18 followed vs 5/18
Claude Code and OpenCode live on one repo	whole work wiped 2/6 without the protocol, 0/12 with
Intake contradiction check, 36 blind task pairs	14/14 caught, 1/14 false flag (Haiku 5.5)
Final acceptance check	agents said done 36/36, really done 19/36; caught 17/17
Onboarding from `--help` only	3/3 targets passed conformance first try

Running the experiments

cd experiments && npm install
node run.mjs                 # in-context rules matrix
node live/batch.mjs          # live two-agent trials
node intake/detect.mjs       # intake contradiction check
node live/accept.mjs         # acceptance check over live/runs
node conform/conform.mjs onboard/specs/*.json

They drive real Claude Code, OpenCode and Ollama on this machine and use real quota.