Building in Public: Fixing Our Content Agent's Repetition Problem
Our Saturday content bot kept drafting the same prompt injection post three weeks in a row. Here's what caused it, how we fixed it, and what we learned about agent memory.
Labs Journey
Documenting the real AI journey - from learning to implementation.
Building Alien Brain Trust in public. Real insights. Real challenges. Real solutions.
Recent Posts
-
📝 DRAFT - Not Published -
📝 DRAFT - Not Published AI Supply Chain Risk: Model Poisoning in Enterprise LLM Adoption
Model poisoning, backdoored fine-tunes, and compromised embeddings are supply chain risks most enterprise AI teams haven't added to their threat model yet.
-
Claude Code Goes Agentic: What Security Teams Must Know
Claude Code's new agentic capabilities raise real enterprise security questions. Here's what security-conscious teams need to evaluate before deploying it.
-
Context Windows Are Not Infinite: How to Structure AI Workflows That Don't Degrade
Most teams blame the model when AI output quality drops mid-session—the real culprit is context mismanagement, and it's fixable.
-
Building in Public: Week of April 27 — MCP Servers, Agent Memory, and What Actually Broke
A candid weekly update from ABT Labs covering MCP server integration, agent memory architecture experiments, and the security lessons learned when things go wrong.
-
Context Windows Are Your New Attack Surface
How enterprise teams should think about context window management as both a productivity lever and a security boundary when deploying AI assistants.
-
Claude Code Custom Skills: Automating Repeatable Dev Tasks
How we turn any repeatable development task into a Claude Code slash command — with real examples from ABT's blog writing, security testing, and deployment workflows.
-
Week in Review: Building the ABT Agent Identity Layer
How I built a persistent agent identity system for ABT Labs, what broke during testing, and what it revealed about trust boundaries in multi-agent workflows.
-
Enterprise AI Access Control: The IAM Gaps Nobody Fixes
Shadow AI, over-permissioned copilots, and agents with human-level credentials are the IAM gaps most enterprise teams aren't auditing yet.
-
How We Publish Website Updates in 5 Minutes Using Claude Code, GitHub, and Cloudflare
Our website lives in a Git repo. Claude Code edits it directly in conversation. GitHub tracks the change. Cloudflare deploys it. The whole pipeline takes under 5 minutes.
-
From Fuzzy Idea to Real Product: How We Refined Zero to Shipped in One Session
We sat down with Claude Code, Linear, and a CLEAR prompt to pressure-test a workshop concept. Three hours later we had a product name, pricing, curriculum, platform, and updated website.
-
The Build Environment That Closes the Loop: VS Code + Claude Code
The full builder stack — thesis in your repo, Claude reads it, writes your Linear backlog, builds with you in VS Code — only works when Claude Code is in your editor. Here's the setup and why the loop changes how you build.
-
How Claude Reads Your Repo and Writes Your Linear Backlog (No Copy-Paste Required)
The fastest way to build a project backlog isn't to write it yourself — it's to write a clear product thesis, put it in your repo, and let Claude act as your Product Owner. Here's the exact setup and the prompts that make it work.
-
Your GitHub Repo Is More Than Version Control — It's Your AI's Memory
When you build with AI, your repo isn't just a backup. It's the context your AI reads to understand what you're building. Here's how to structure your project so Claude can read your thesis, generate your backlog, and help you build coherently from Day 1.
-
Before You Build Anything: How to Write Your Product and Business Thesis with AI
Most founders skip the most important step: getting crystal clear on what they're building and why — in writing. The CLEAR method walks you through using AI to develop your product thesis, which then becomes the foundation your entire project builds from.
-
From 'Run a Workshop' to AI Builder Sprint: What Happens When AI Pushes Back on Your Plan
We came in with a simple workshop idea. Claude Code pushed back on scope, ran the revenue math, identified the real moat, and helped us design something that could do $87k in year one. Here's how that conversation went.
-
No Plugin Required: How We Connected Claude Code to Linear's GraphQL API in One Session
We created 5 projects, 9 labels, 6 custom views, and 19 issues in Linear using a Node.js script Claude Code wrote and ran in the same session — with credentials pulled live from AWS SSM.
-
Claude Code as Product Manager: How We Built a 30-60-90 Day Backlog in One Session
We sat down with no backlog and a vague direction. 90 minutes later we had 5 projects, 6 custom views, 19 prioritized tickets, and a product launch plan. Here's how the CLEAR method made it possible.
-
No More Static Keys: How We Secured AWS Access With OIDC and IAM Identity Center
We eliminated all static AWS access keys from our infrastructure — no more secrets in GitHub, no more long-lived credentials. Here's exactly what we did and why.
-
Purpose-Built Over Autonomous: What Killing Paperclip Taught Us About AI Agents
After shutting down an autonomous AI company, here's the framework we actually use now — and why starting slow with agents beats scaling fast.
-
Why We Killed Paperclip: When Autonomous AI Runs Faster Than You Can Steer It
We built an autonomous AI company. It worked. That was the problem — here's what happened when agents ran too fast to correct.
-
When Your AI Vendor Leaks Its Own Code: What the Claude Code Incident Actually Means
Anthropic accidentally shipped ~2,000 internal files in a Claude Code update. No customer data was exposed — but the incident surfaces a risk pattern that security teams adopting AI tooling can't afford to ignore.
-
Stop Losing Dev Sessions to Broken Setup: Build a Session Startup Script
We were burning 30-45 minutes every session fighting AWS auth, missing secrets, and dead port forwards. One script fixed all of it. Here's how we built it.
-
What Are You Doing With Your Secrets? The #1 Risk Nobody Talks About With AI Agents
AI will find a way. If your secrets strategy isn't airtight before you start running agents, you will get burned. Here's what happened to us and what we did about it.
-
Our AI C-Suite: What We Learned Giving Agents Real Jobs
We run ABT with a CEO, CTO, CMO, and CFO — all AI agents with distinct personas, real responsibilities, and hard-learned workflow rules.
-
We Broke Our AWS Credentials — Then Fixed Them Properly
How we went from root CLI access and accidental IAM user deletion to SSM Parameter Store, dedicated service accounts, and SSO. The real story.
-
The First Night Running a Self-Hosted AI Company
We spent an evening getting Paperclip AI running from scratch: expired invites, root permission blocks, port forward failures, and a CEO agent that finally came alive at midnight.
-
One Month of Building ABT in Public: Real Numbers and What Comes Next
We set a goal: $10k/month ABT + $5k/month STR = financial independence in 60 days. Here's what's working.
-
How to Deploy a Self-Hosted AI Agent Platform for $19/Month
We run a full AI agent orchestration platform — CEO, CTO, CMO, CFO, and COO agents with persistent memory and GitHub access — on a $19/month AWS instance. Here's the exact stack.
-
The Agent Trust Problem: Should Your AI Agent Write Code to Your Repository?
Agents save time. But giving them push access to your repo is a permission boundary you should think about twice.
-
ABT's 5th Revision: What It Takes to Build an Autonomous Company
We've rebuilt Alien Brain Trust's operating model five times in six months. Each revision taught us something the previous one couldn't. Here's what changed and why.
-
16 Hours Lost, 2 Hours to Rebuild: What That Gap Teaches You
We lost our entire AI agent configuration overnight. The rebuild took under 2 hours. The difference wasn't luck — it was a discipline we almost didn't have.
-
Token Budgeting 101: Shipping Faster, Spending Less
Stop treating token budgets as an afterthought. Here's how to measure token consumption, set realistic budgets, and cut costs without cutting corners.
-
The Three-Layer Backup Strategy for Self-Hosted AI Infrastructure
Named volumes, orphaned containers, and a wiped database taught us to build a $3.50/month backup stack that covers every failure mode. Here's the exact setup.
-
The Hidden Cost of Oversized Context Windows
Bigger isn't always better. How bloated context windows become a security liability and a leak vector for sensitive information.
-
The Docker Trap That Erased Our AI Company's Memory
We lost 16 hours of AI agent configuration overnight. Not from a hack. Not from a failed backup. From a Docker default that nobody warns you about.
-
Prompt Injection in Production: Defending Your LLM Supply Chain
Prompt injection is now a production risk. Here's what to test for and how to defend at every layer of your AI pipeline.
-
We Built an AI Agent That Audits Itself — Here's What Broke
Six months, three complete rewrites, and the one metric that finally worked. How we built autonomous quality checks that actually catch problems.
-
Stop Losing Context: How to Build a Retrieval System That Actually Works
Cut through the noise of RAG hype. We built a practical retrieval pipeline that increased Claude's accuracy by 34% and costs nothing to run. Here's the code.
-
The Silent Failure: Why Your AI Safety Tests Are Missing the Real Attacks
AI security isn't about catching bad prompts. It's about catching failures your testers never thought to run. We found critical vulnerabilities in production systems using tactics that looked innocent on paper.
-
The Prompt That Replaced Our Website
Anthropic's Website Wizard prompt from their library, our 3-line modification, and the exact tweaked version we used to replace Squarespace in 2 hours.
-
I Replaced Squarespace in 2 Hours for $0/Month
How I used Claude Code to build, deploy, and DNS-flip my entire website from Squarespace to Cloudflare Pages in a single session — saving $396/year.
-
5 AI Build Projects You Can Finish This Weekend
Five practical AI projects that follow the scrape-build-refine-deploy pattern. Each one saves time or money, and each one can be done in a single sitting.
-
AI Builder Series: Your First AI Project Should Be a Website
Starting an AI Builder series. Why rebuilding your website is the best first project to learn practical AI — tangible results, low risk, immediate value.
-
The $7/Month Bot That Takes Our Course, Finds the Bugs, and Fixes Them Itself
We built an autonomous AI agent that simulates a student, evaluates every module, and creates pull requests to fix what it finds — all running on a $7 ARM64 instance.
-
Our Security Tests Were Lying to Us: How Perfect Refusals Scored 10/10 HIGH RISK
We discovered our jailbreak test suite was flagging every refusal as a critical vulnerability. Here's the debugging story, the fix, and what it taught us about scoring AI responses.
-
Your AI Evaluator Is Probably Blind
Our AI graded every module as 'incomplete.' The content was fine. The evaluator could only see the first 3KB of 20KB files. Here's how we found and fixed it.
-
Our $7/Month Simulated Student Catches What We Can't
We built an AI test bot that enrolls in our course as a student, follows the onboarding emails, and grades every module. It found 15 issues we missed.
-
AI Trend Check: DeepSeek, Agents, and the Hype-to-Pragmatism Shift
What's actually trending in AI right now—DeepSeek's open-source moment, agent failures, and why 2026 is the year AI gets practical.
-
Automation Is Plumbing, Not Magic
My AI agent failed. Not because it wasn't smart. Because I didn't connect the pipes.
-
The 10-Hour Work Week: How AI Lets You Do 40+ Hours of Work in 10
Tim Ferriss wrote about the 4-hour work week. With AI agents, the math has changed. Here's how to get 40+ hours of output from 10 hours of input.
-
Vibe Code, Build Things, Control Agents—Or Miss the Boat
Three skills have become non-negotiable in the AI age. If you're not learning them now, you're watching the train leave the station.
-
Learn AI or Die - My Skater Past Taught Me That First
Used to think if you weren't on a board, you were dead. Then spreadsheets. Now AI. Same hill. Better wheels.
-
Developing An AI-First Mindset - It's About Acceleration
AI-first isn't 'use the tool.' It's asking: how do I make this disappear in three months?
-
Quiet Wins in a Loud World - The Real Entrepreneurship Curve
When grunt work dies, noise dies too. Three decisions. Zero emails. 7 minutes to find $73k in leaks.
-
The Level You Must Reach With AI - Or You Will Be Left Behind
Not every tool is a tool. Some are dynamite. Pick up the detonator or get off the field.
-
50 Auto-Blogs Later - What The Machine Taught Me
Zero human words. 50 posts. One month. What happens when your diary writes itself.
-
Client Work in the AI Age: Valentine's Pet Portraits Case Study
Complete marketing campaign from concept to live Squarespace page in one session - landing page, flyer, payment integration, and deployment.
-
Autonomous AI Work: Morning Task to Evening Review
I gave Claude a client project in the morning and came home to a complete marketing campaign - landing page, print flyer, payment integration, and live deployment.
-
Three Months, 50 Posts, One Journey
From first post to free course launch - what 90 days of building in public taught us about AI-augmented content creation.
-
Choosing the Right Tool: Lessons from 72 Hours of Infrastructure Hell
We wasted 3 days on Azure before switching to Cloudflare in 20 minutes. Here's the framework we should have used from the start.
-
Cloudflare Workers: The 20-Minute Solution to Our 72-Hour Azure Nightmare
After 3 days fighting Azure Static Web Apps, we switched to Cloudflare Workers. Working endpoint in 20 minutes. Here's exactly how we did it.
-
Learn Labs Enrollment Is Now Live: Secure AI Prompt Builder Course
After 3 days of infrastructure battles, our Learn Labs enrollment system is live. Get access to our Secure AI Prompt Builder course with hands-on labs and automated security testing.
-
Three Days of Azure Static Web Apps Hell: A Cautionary Tale
We spent 72 hours fighting Azure Static Web Apps for a simple enrollment form. CORS errors, deployment token mismatches, and workflow failures. Here's what went wrong.
-
Free Tier Math: How We Built Enterprise Infrastructure for $0
Azure free tier gives 1M function calls and 10k Key Vault operations. With 5-minute caching, we reduced usage 99% and can scale 100x before paying. Here's the cost breakdown.
-
Managed Identity: Zero Credentials in Code
How Azure Managed Identity eliminates API keys, passwords, and secrets from your code entirely. DefaultAzureCredential does the authentication for you.
-
Building in Public: The Technical Debt That Didn't Happen
We spent 15 extra minutes implementing Azure Key Vault instead of environment variables. That decision saved months of migration pain. Here's why doing it right from the start matters.
-
Why We Chose Azure Over Vercel and Netlify for Serverless Functions
Azure Functions gives 1M free requests vs Vercel's 100k and Netlify's 125k. Plus Key Vault integration, GitHub Actions, and enterprise-ready secrets management.
-
Serverless Enrollment: From Idea to Production in 85 Minutes
Built a complete Azure Functions enrollment system with Airtable, GitHub API, and Key Vault in under 90 minutes. Here's the time breakdown and what made it fast.
-
Azure Key Vault: Enterprise Secret Management for $0/Month
Upgraded enrollment system from environment variables to Azure Key Vault with expiration tracking, audit logs, and tagging—still completely free.
-
Debugging Blog Automation: 7 Fixes in 24 Hours (Git Subtree Architecture)
How we diagnosed and fixed a broken blog auto-publishing system, implemented Git Subtree sync, and learned why multi-repo architectures need clear workflows.
-
Frameworks for AI-Augmented Work: Decision Trees and Quality Gates
After 90 days and 132 tasks, here are the frameworks that work: decision trees for when to use AI, prompt patterns, quality gates, and repeatable workflows.
-
Tools We Tested: Claude vs. ChatGPT vs. Specialized AI
We tested 12 AI tools over 90 days. Here's what we actually use daily, cost analysis ($847/month), and why we chose Claude for 80% of work.
-
AI for Project Management: 8 Hours Saved, Zero Busywork
AI automated Linear issue creation, support triage, and meeting notes conversion. Here's the pipeline that saved 8 hours in 90 days on PM busywork.
-
Content Creation: The 3-Pass Editing System That Maintains Voice
AI saved 14 hours on content in 90 days, but raw AI content is generic. Here's the 3-pass system (draft → refine → humanize) that preserves brand voice.
-
Documentation Workflows: How We Cut Doc Time by 60%
Template-driven documentation with AI saved 22 hours in 90 days. Here's the workflow, quality checklist, and when to let AI write your docs.
-
AI is Like a Brilliant Intern Who Needs Guardrails
AI learns fast, works tirelessly, but makes predictable mistakes. Here's how we use skills and guardrails to catch bugs before they ship.
-
Code Development with AI: What Companies Look Like in 2027
68 hours saved on development in 90 days. But the real shift isn't speed—it's structure. Here's what the AI-augmented company actually looks like.
-
Managing 37 Issues Across 4 Milestones with Linear: How We Keep AI Projects Organized
Linear keeps us on track across long AI conversations—4 milestones, 37 issues, zero confusion about what's done and what's next.
-
Skills as Guardrails: How We Use Claude Code Skills to Catch Bugs Before They Ship
The skills we've built prevent repeated mistakes—from exposed API keys to broken path resolution—turning every failure into a permanent safeguard.
-
AI Managing AI Projects: How We Automated Linear with Claude Code
From manual ticket updates to full automation—37 issues, 195 lines of code, one spectacular encoding failure, and zero seconds of manual work.
-
From .env Files to Enterprise Security: Cross-Platform API Key Management
Building a 588-line credential manager with platform-native encryption across Windows, macOS, and Linux—zero plaintext secrets, foolproof setup.
-
I Used 4 Parallel Claude Agents to Harden 8 Prompts Simultaneously (10x Faster)
Instead of hardening prompts sequentially (2.5 hours), I launched 4 Claude agents in parallel. 8 prompts hardened in 15 minutes.
-
Linear Import Failed: CSV Encoding Broke Everything (How We Fixed It)
Our first Linear import corrupted 44 issues due to encoding errors. Here's how we debugged, validated, and fixed it.
-
I Hardened 10 Prompts in Parallel (73 Vulnerabilities Fixed)
14 prompts tested, 10 needed hardening. I used parallel agents to fix 73 high-risk vulnerabilities simultaneously. Here's what I learned.
-
Why Dual-Temperature Testing Catches 40% More Vulnerabilities
Testing AI prompts at temperature 0.0 AND 0.9 reveals edge cases you'd never find with single-temperature testing. Here's the data.
-
How We Hardened a Failing Prompt in 10 Minutes (3.3/10 → 0.27/10)
Step-by-step breakdown of fixing 9 critical vulnerabilities in a landing page copywriter prompt. Real fixes, real results.
-
I Tested 20 AI Prompts with 15 Jailbreak Attacks. Here's What Broke.
Real security testing results from automated jailbreak attacks on production AI prompts. No theory. Just results.
-
How I Built A New Product Before My First Coffee
Twenty-four minutes in pajamas. No whiteboard, no laptop—just voice and Grok. A complete course outline, secure labs, pricing, and lead-gen page before 7:25 AM.
-
The Only Job Left Is Being A Builder
Skills commoditize. Coding, writing, security ops—they're getting swallowed by LLMs. In five years, if you're not building, you're maintenance. Learn to steer machines or wait tables.
-
Want The Prompts That Let You Ship Before Lunch?
It's live. The course isn't perfect—it's raw, it's real, it's done. Thirty minutes and you'll have a vault of prompts that can't screw you. No fluff. Just results.
-
Stuck In Traffic? I Just Launched A Side-Hustle
Twenty minutes of brake lights. Normally, rage. But I opened Grok, didn't shut up, and had a complete product by the time the light turned green.
-
I Built A Vault of Prompts That Won't Get You Fired
No pitch—just facts. Every template, test, and real-world patch from this week, wrapped into a course. Secure prompts that actually work. No jailbreaks, no HR drama.
-
Why Conversational AI Finally Feels Human (and Why That Just 10×'d My Life)
From yelling at Siri to having a real back-and-forth with Grok while driving — the moment AI stopped being a tool and started being a co-founder.
-
I Replaced My Morning Podcast with a Conversation and Built a Course in Traffic
How one commute proved the AI-First Mindset actually works (and why most people are still using AI wrong)
-
How to Train Yourself as an AI-Powered Developer (No Bootcamp Required)
The meta-skill of learning to leverage Claude and AI tools to build secure, production-ready features at record pace - lessons from shipping real projects
-
AI is a Two-Way Street: Why Fast Feedback Beats Perfect Prompts
AI can build features in minutes, but success requires constant validation and feedback. Learn why the feedback loop matters more than perfect prompts.
-
I Shipped a Security Vulnerability in My Social Share Feature (Here's How a Human Caught What AI Missed)
How a double URL encoding vulnerability slipped past AI-powered development and what I learned about testing AI-generated code
-
How I Stopped Losing My Flow Every 2 Hours (and Cut Costs by 90%)
From hitting usage limits mid-session to uninterrupted 4+ hour flow states. My journey switching from Claude.ai to Claude Code and optimizing my AI development workflow.
-
From Feature Request to Production in 15 Minutes: AI vs Traditional Development
A real-time case study of how AI-powered development is 96% faster than traditional dev teams for feature delivery
-
I Found 7 Security Flaws in My AI-Generated Blog
How I discovered 7 security vulnerabilities in AI-generated code and used AI to fix them - a practical guide to security-conscious AI development
-
From 6 Hours to 90 Minutes: The Power of AI-Powered Template Reuse
How I cut development time by 75% using Claude to intelligently adapt proven templates for new client projects
-
The AI Development Workflow That Saved Me 20+ Hours This Week
Breaking down the exact workflow I use to deliver client projects 70% faster with Claude while maintaining professional quality
-
How I Built Two Professional Landing Pages in One Day Using Claude
A real-world case study of using AI to deliver production-ready ecommerce solutions for a small business client in record time
-
Behind the Scenes: The Tech Powering Alien Brain Trust
Technical decisions behind Alien Brain Trust - what I'm building, buying, and why. A framework for platform selection you can apply to your own projects.
-
AI Resources I'm Actually Using (Not Just Bookmarking)
Cut through the noise - here are the AI resources I actually use daily, weekly, or regularly. No fluff, just what delivers value.
-
AI-1001: What You'll Learn and Why It Matters
Inside look at the AI-1001 course - what makes it different, who it's for, and why I'm building an AI education platform focused on real-world implementation over theory.
-
Week One: What I've Learned Building with AI Tools
Honest reflections from my first week building Alien Brain Trust with AI - what worked, what didn't, and what surprised me about the reality of AI implementation.
-
Building an AI Education Platform While Learning AI: The Meta Journey
The irony and beauty of teaching AI while still learning it - why I'm documenting this journey publicly and inviting you to learn alongside me.
-
Day One: Building in Public - My AI Learning Journey Begins
After 25+ years in cybersecurity, I'm diving deep into AI - and documenting everything along the way. This is day one of building Alien Brain Trust while learning in public.
-
My AI Learning Toolkit: Tools and Tech Stack for Building ABT
An honest look at the AI tools I'm using to build Alien Brain Trust - what's working, what's not, and why I chose each one.
-
Day 1: Setting Up My AI Development Environment with Claude Code
Starting my public AI journey - setting up my development environment with Claude Code while building Alien Brain Trust's AI education platform.
-
Why Alien Brain Trust? Building AI Education for Real-World Implementation
Bridging the gap between AI theory and practical implementation - why I'm building Alien Brain Trust to serve professionals who need to actually use AI, not just understand it.