Browse / General-purpose / ai-assisted-dev-kit

ai-assisted-dev-kit

ikabeee

My Claude Code setup for AI-assisted development: global agent rules and a multi-harness workflow skill

Open ikabeee/ai-assisted-dev-kit → Checked against the four rules.

Built for one agent

ships CLAUDE.md and nothing for other agents.

In its own words

README.mdikabeee/ai-assisted-dev-kit

My Claude Code setup for AI-assisted software development: the global rules I give every agent and the workflow I use to split work across several coding agents.

The idea behind it: the model is a member of the team, not a search box. I write the score (rules, specs, context), the agents play it, and I review before anything ships.

The workflow

Work starts from a general objective, not from a single request. The general objective splits into specific objectives, one feature each, tracked in an objectives index (odd/objectives.md). Each feature runs through the flow below. When it closes, the index points to the next one and I start a fresh session, so every feature begins with a clean context and the quality stays the same from the first feature to the last.

This objective layer is my own addition. I proposed it upstream in gentle-ai#5715, and the fresh session at feature close in gentle-ai#5156.

  • All harnesses work in the same working tree, each inside a declared, disjoint edit surface.
  • Only Claude runs git. Harnesses never commit.
  • Every API or technique is backed by the current official docs, with the URL cited.
  • The base workflow comes from Organic Driven Development (ODD) in gentle-ai: the inline path for small adjustments, feature documents with numbered specs, disjoint edit surfaces in a shared tree, and ~400-line work units.
  • The coder → cleaner pipeline comes from Uncle Bob's swarm-forge "two-pack". I follow the process without its tooling.

Quoting 62% of the README. Read the rest on GitHub ↗

Read from the repository's git tree on 2026-08-17.