David Ichim ichim-david

Humble Pi — local agentic coding on minimal hardware

A coding agent that runs entirely on your own machine. No API keys, no cloud, works offline.

pi is the agent you talk to. It runs against a local llama.cpp server hosting Google Gemma 4 (the unsloth builds).

Pick a model by how much memory you have:

Model	unsloth repo	Download	VRAM	Start	Max

name	agents-md
description	How to write, audit and edit AGENTS.md files.

AGENTS.md Authoring

Purpose

Create and improve AGENTS.md files that are short, scoped, and useful to coding agents. Preserve intent and non-obvious rules; do not summarize the repo.

Making MLX Go Fast

This guide assumes you have some familiarity with MLX and want to make your MLX model or algorithm as efficient as possible.

The guide covers the following:

Graph Evaluation
Type Promotion
Operations

patch-edid.rb

A script to fix EDID problems on external monitors in macOS.

Instructions

Connect only the problem display.
Create this directory structure (if it doesn't already exist):

Comparing Approaches for GitHub Copilot + Java

Overview

When GitHub Copilot analyzes or generates Java code, the approach used for code intelligence directly impacts performance (speed of interaction), output quality (correctness and idiomatic usage of Java, Spring, and related frameworks), resource consumption (tokens and premium requests), and cost.

This document compares six approaches across these dimensions.

Twitter Mention Blocklist Policy

Public-safe version. Redacted. No handles, no links, no callouts.

Block immediately

Strongly derogatory replies
Crypto spam, token shills, wallet/contract junk
Clear AI slop or reply-bot behavior
Repetitive low-context spam

description	Mini Beast

<tool_preambles>

Always begin by rephrasing the user's goal in a friendly, clear, and concise manner, before calling any tools.
Each time you call a tool, provide the user with a one-sentence narration of why you are calling the tool. You do NOT need to tell them WHAT you are doing, just WHY you are doing it.
- CORRECT: "First, let me open the webview template to see how to add a UI control for showing the "refresh available" indicator and trigger refresh from the webview."
- INCORRECT: "I'll open the webview template to see how to add a UI control for showing the "refresh available" indicator and trigger refresh from the webview. I'm going to read settingsWebview.html."
ALWAYS use a todo list to track your progress using the todo list tool.

layout	post
title	We need practical AI workflows
date	2025-11-17 08:40:00 +0000
categories	posts
permalink	/posts/promptfiles-vs-instructions-vs-agents/

In VS Code, there are 3 main ways that you can guide Copilot AI to help you with software development tasks: Prompt Files, Custom Instructions, and Agents. Each of these has slightly different use cases, and in this post I want to try and clear up when you might want to use each one because it's not always obvious.

	# Install PyTorch separately for the B300 / Blackwell CUDA stack, for example:
	# pip install --pre --upgrade torch torchvision --index-url https://download.pytorch.org/whl/nightly/cu128
	#
	# Then install this file:
	# pip install --upgrade -r requirements_gemma4_additive_moe_reviewed_v9_streamfix.txt
	#
	# FlashAttention 2 is optional. The script defaults to SDPA for maximum
	# out-of-the-box compatibility. To use FlashAttention 2 after installing it:
	# pass --attn_implementation flash_attention_2.
	# To install FlashAttention explicitly:

	#!/usr/bin/env bash
	# Atomic commit helper to prevent AI's from accidentally deleting
	# changes from dangerous commands. I only allow AI agents to commit with this flow.
	# I disallow AI agents from git reset, git checkout, and git revert because
	# they are dangerous operations when multiple agents (or I) are editing
	# files in the same work tree.
	set -euo pipefail

	usage() {
	cat >&2 <<'EOF'