Skip to content

Zejnilovic's agentic defaults

Sensible defaults for agentic coding: test-first, reasonable coverage, and context that stays current. Generic instructions, skills and agents for Claude Code and GitHub Copilot, built for practical delivery on modest model budgets.

Quick start Choose your path Read the method GitHub

Install it in two commands

brew install zejnilovic/tap/agd

curl -fsSL https://raw.githubusercontent.com/Zejnilovic/agd/main/install.sh | bash

Then, in your project's root:

agd init

agd is the installer: it copies the kit into your project and updates it later. Its own documentation covers everything else it does. Don't want a tool? Copy the files by hand.

What you get

  • A working agreement

    One short AGENTS.md, loaded into every conversation: deliver small, test first, start from written context, work economically, and a clear definition of done.

    Overview

  • Agents: who does the work

    Thirteen roles, from analyst and architect to developers, tester, reviewers and documenter. Cheap models read and execute; strong models think.

    Agents

  • Skills: how it's done

    Practices that load only when a task needs them: TDD, debugging, task plans, diagrams, project scans, and Flask, Spring, React and SQL.

    Skills

  • Context that stays current

    CONTEXT.md files describe your system for agents, and every task ends with a check that keeps them true to the code.

    Keep CONTEXT.md current

  • Only what you use

    Install in packs: a core, plus research and the Flask, Spring, React and database developers your project needs.

    Packs

  • Claude Code and Copilot

    Plain Markdown, shell scripts and JSON in formats both tools read. Tools that read AGENTS.md, such as Cursor, get the working agreement.

    How they differ

How a task flows

Small changes need nothing special: just chat, and the working agreement applies. Bigger ones go through roles, each handing its result back to you:

flowchart LR
  ticket["Ticket and notes"] --> analyst --> architect --> developer --> tester --> reviewer --> documenter
  documenter --> ctx["CONTEXT.md, decision records"]

How work flows

Why it works this way

The method: Working with coding agents

The kit is one way to put a method into practice. Working with coding agents explains it: the principles, rules of thumb, anti-patterns and nine hands-on exercises. It's written to stay true as tools and models change. The short version: writing code is cheap now; knowing what is correct is not.

Read the method