title	description
Weekend Coding Agent	Build an AI coding assistant step by step.

Weekend Coding Agent

In this project we are going to build a software coding agent from scratch. We are going to bootstrap it with an existing coding agent -- it doesn't really matter which one, as long as it's using one of the frontier models -- and then quickly going to use it to improve itself.

You'll learn the concepts of how coding agents work, and see how to layer on functionality to make it more productive.

What we're building

Phase 1: Prompts to Write Prompts

We'll develop a research and knowledge gathering prompt. This will form the basis of giving the agent the necessary background knowledge and preplanning to build the software we need.

This tutorial uses Gemini 3 and Typescript -- feel free to ask it about different technologies if you want. Build it with python on Antropic. Use ruby and grok. Ask it to help you decide what technolgies there are.

Phase 2: Bootstrapping the Agent

Once we know what we want to do, we'll bootstrap the agent. This is telling the agent to go to https://ampcode.com/how-to-build-an-agent and build something like that.

And then it does. We start running it to make sure that it works.

Phase 3: Using the Agent to add features to the Agent

At this point we will be using the agent we just developed to develop it further, adding

• Tool execution - Giving the agent more capabilities
• Context Awareness - Tuning the system prompt
• Command System - Adding slash commands
• Session Analysis - summarizing past sessions, identifying goals, completed work, active tasks, and lessons learned

The Philosophy

This project demonstrates three key ideas:

1. Prompts to write prompts - The /research command is a prompt that generates research reports, which are themselves prompts for future work.

2. Specialized knowledge - By loading research reports into the system prompt, the agent has domain-specific knowledge about TypeScript, Gemini 3, web search APIs, etc.

3. Self-improvement - The agent can summarize its own sessions, identify what it's learned, and resume previous work. It can debug itself when errors occur.

The frontier models (Gemini 3 Pro, Claude, GPT-4) are so capable that with just a few tools—read, list, edit—they can build complex systems, research solutions, and improve themselves.

See Also

• BUILD_PLAN.md - Technical implementation plan
• reports/ - Research reports generated by the agent
• How to build a coding agent - By Geoffrey Huntley