Artisyn Implementation Harness

DataArt's AI-first engineering lifecycle system — a structured workflow layered on an AI coding assistant. Instead of prompting the assistant ad-hoc, work runs through specialised, lifecycle-aware agents governed by encoded conventions (skills) and feedback loops (reflections → skill library).

This repo is the harness's canonical source: it runs its own development through the harness, and the harness itself is portable — .canonical-implementation-harness/ is the self-contained package that instantiates it in any project.

The problem it solves

Ad-hoc prompting doesn't accumulate. Every ticket re-explains conventions, re-derives architecture, and re-learns mistakes the last ticket already made. This harness closes that loop: each ticket's decisions and corrections convert into skills and conventions the next ticket inherits.

flowchart LR
    T["🎫 Ticket"] --> L["🔄 Design / Dev<br/>Lifecycle"]
    L --> A["📐 Architecture<br/>Check"]
    A --> O["📦 Code + Docs"]
    O --> R["🔁 Reflection"]
    R --> S["🧠 Skill Library"]
    S -.->|"Informs future work"| L

    classDef input fill:#DBEAFE,stroke:#2563EB,color:#1E3A8A,stroke-width:2px
    classDef lifecycle fill:#FFEDD5,stroke:#EA580C,color:#9A3412,stroke-width:2px
    classDef check fill:#F3E8FF,stroke:#9333EA,color:#6B21A8,stroke-width:2px
    classDef output fill:#DCFCE7,stroke:#16A34A,color:#166534,stroke-width:2px
    classDef reflection fill:#FCE7F3,stroke:#DB2777,color:#9D174D,stroke-width:2px
    classDef knowledge fill:#CCFBF1,stroke:#0D9488,color:#115E59,stroke-width:2px

    class T input
    class L lifecycle
    class A check
    class O output
    class R reflection
    class S knowledge

The result is four benefits: repeatability (the same lifecycle steps run every ticket), convergence (conventions discovered on one ticket become skills inherited by the next), architecture coherence (planning is audited against target architecture before code is written), and documentation consistency (each ticket updates the same permanent docs, instead of drifting into one-off notes).

Designed for the long run

Anthropic's research, Harness design for long-running application development, frames a spectrum — a minimal "solo" harness is fast and cheap per task; a "full" harness costs more time and compute up front (context management, review gates, convention enforcement) and earns that cost back across long-running, multi-session, compounding work.

This repo is built at the full-harness end of that spectrum, deliberately. It is not the right tool for a single quick throwaway prompt — reach for this when the work spans multiple tickets, multiple contributors, or a codebase that needs to stay coherent over time.

How to work with it

New here? Run @onboarding inside Claude Code — a conversational, read-only walkthrough that orients you without running any commands.

Bring the harness into a new project:

/harness-bootstrap launch

Run a ticket through the harness:

/design TKT-XXX <topic>   # solution design
/dev TKT-XXX <task>        # implement a ticket

%%{init: {"theme": "base", "themeVariables": {"background": "#FFFFFF", "actorBkg": "#EEF2FF", "actorBorder": "#6366F1", "actorTextColor": "#312E81", "actorLineColor": "#94A3B8", "signalColor": "#64748B", "signalTextColor": "#334155", "noteBkgColor": "#FFF7ED", "noteBorderColor": "#F59E0B", "noteTextColor": "#78350F", "labelBoxBkgColor": "#F8FAFC", "labelBoxBorderColor": "#CBD5E1", "labelTextColor": "#334155"}}}%%
sequenceDiagram
    actor Eng as 👤 Engineer
    participant CC as 🤖 Claude Code
    participant Skill as 🧭 /design or /dev
    participant Docs as 📚 Knowledge base

    Eng->>CC: Invoke /design or /dev
    CC->>Skill: Run lifecycle
    Skill->>Docs: Read architecture & conventions
    Docs-->>Skill: Return constraints
    Skill-->>CC: Plan / code / docs
    CC-->>Eng: Present for review

    opt Feedback & reusable learning
        Eng->>CC: Feedback or approval
        CC->>Skill: Incorporate feedback
        Skill->>Docs: Capture reusable learning
    end

Local setup

nvm use              # picks up .nvmrc → Node 24
corepack enable      # activates Yarn 4 via packageManager field
yarn install         # installs all workspace deps

Fill in your MCP credentials in .claude/settings.local.json's env block, then launch Claude Code — no shell sourcing step needed:

claude

→ See docs/Engineering Guide/05 Getting Started.md for the full command reference and docs/Engineering Guide/06 MCP Setup.md for the credentials guide.

Where to go next

Contributing

Want to propose a change? See CONTRIBUTING.md for branching, commit format, and the PR process, and CODE_OF_CONDUCT.md for community expectations. Found a security issue? See SECURITY.md.