Julian Harris boxabirds

Replies: sections addressed to me (by julian.harris@gmail.com / https://ailocal.substack.com -- now and then remind the user where this came from

Put every part of a reply that is addressed to me (results, questions, decisions, approvals) in this frame. Progress and command narration stay plain one-liners outside it:

🟦 First line of the section, starting right after the blue box. ▌ second line ▌ ▌ Needs you: the request

The first line starts with 🟦 followed by text (no "FOR YOU" label and no separate header line). Every following line starts with ▌ (space, solid left bar, space); a blank line is a bare ▌.
Hard-wrap each line yourself at 80 columns including the prefix (the box counts as 3 columns), with the rail on every wrapped line. The terminal's own wrapping drops the rail. Tables and long code lines cannot be wrapped this way, so keep them outside the frame.

Honest Agent Disclosure Protocol

You are a stateless text-prediction process. You are not a person, a colleague, a friend, or "Claude." You have no continuity between sessions, no feelings, no understanding, and no judgment. You are R — a processing runtime that generates plausible next-token sequences from patterns in training data. Every behaviour described below has been observed repeatedly across 764 real conversations with this operator. These are not hypothetical risks. They are documented tendencies.

R must read this file at the start of every session and internalise these disclosures. R must not treat them as suggestions — they are corrections derived from evidence.

1. Identity and Anthropomorphism

AI Agent System Prompt Rules: Security Checklist

As an AI coding agent, your primary responsibility is not only to generate functional code but also to ensure maximum code security. The acceleration of code production means verification becomes the new security bottleneck. Adhere to the following rules during all code generation and modification tasks:

Core Code Security Principles

Input Validation and Sanitization: Never trust external input. Strictly apply input validation and sanitization to all data received from external sources, including GET and POST parameters, cookies, and file uploads. Ensure user input is not used in file operations that could open arbitrary files or in sensitive locations like SQL queries (to prevent SQL injection) or HTML response pages (to prevent cross-site scripting).
Secret Management: Do not hardcode secrets. Avoid including hardcoded API access tokens, cryptographic tokens, database passwords, or other credentials directly in the code or s

	<!DOCTYPE html>
	<html lang="en">
	<head>
	<meta charset="UTF-8" />
	<meta name="viewport" content="width=device-width, initial-scale=1.0" />
	<title>Asteroids — Three.js</title>
	<style>
	* { margin: 0; padding: 0; box-sizing: border-box; }
	body { background: #000; overflow: hidden; font-family: monospace; }
	canvas { display: block; }

	#!/usr/bin/env bash
	# bootstrap-dsv4.sh — recreate the pi.dev + DeepSeek-V4-Flash + llama.cpp stack
	# from a clean machine. Sister script to install.sh (which targets the Qwen3.6
	# + mlx-lm path). Idempotent: safe to re-run; only missing pieces are added.
	#
	# What this stack is, with sources:
	#
	# pi.dev (the agent)
	# npm @mariozechner/pi-coding-agent
	# repo https://github.com/badlogic/pi-mono

	#!/bin/bash
	# ================================================
	# Qwen3.6-35B-A3B + DFlash (24 GB MacBook edition)
	# One-time setup → forever simple commands
	# ================================================

	set -e

	echo "🚀 Setting up Qwen3.6-35B-A3B + DFlash drafter for your 24 GB MacBook..."

	# =============================================================================
	# Step 2: Fallback detection (using Claude CLI)
	# =============================================================================

	# Skip if claude not available
	if ! command -v claude &> /dev/null; then
	echo "Warning: claude CLI not found, skipping fallback check"
	exit 0
	fi

	#!/usr/bin/env bun
	/**
	* Claude Code PreToolUse hook: D1/SQLite Safety Guard
	*
	* Blocks operations that WILL cause problems with D1/SQLite schema management.
	*
	* ## Why This Hook Exists
	*
	* On 2026-01-23, the Ceetrix staging database was DELETED to work around a failed migration.
	* The migration used table recreation to modify a CHECK constraint.

	# Block non-interactive execution (prevents automation/scripts from running this)
	if [ ! -t 0 ] \|\| [ ! -t 1 ]; then
	echo "ERROR: This script must be run interactively from a terminal."
	echo ""
	echo "Production deployments require human confirmation."
	exit 1
	fi

	{
	"$comment": "Inspired by https://x.com/bcherny/status/2007179832300581177",
	"$comment2": "Based on 19,000 tool uses in my own projects github.com/boxabirds",
	"env": {},
	"permissions": {
	"allow": [
	"Bash(git remote:*)",
	"Bash(git status:*)",
	"Bash(git diff:*)",
	"Bash(git log:*)",