TechPrufer
Get started
Cole Medin
Cole Medin logoCole Medin · Feature trail

Team-level AI knowledge base with access control

8 videos · Aug 2026 – Sep 2026

What this trail covers

This post extends the concept of team-level AI knowledge bases by demonstrating a practical implementation of an AI agent driving a computer's user interface. The 'Drive Screen' skill allows LLMs to directly interact with the operating system, enabling tasks like morning setup, staging demos, testing desktop applications, and driving other agent sessions. This skill is lightweight, customizable, and integrates directly into projects or can be given to coding agents. The post also touches upon the critical issue of ensuring AI agents adhere to business logic and access control, proposing a solution through Sonar's Hunter Agent which verifies output against intended application behavior rather than just code correctness.

All Cole Medin trails
  1. You Built Your AI Second Brain. Now What? (Here's How to Evolve It)This post details the architecture and implementation of a 'Team Brain' system. It describes a unified data model where information from disparate sources is stored in a single table. The system enforces security and permissions through row-level labeling and access control during read operations. It also outlines how an agent can query this data using its own credentials and token, and how individual agents can build upon this shared knowledge base with their own memory and personality. The technical implementation is demonstrated using Oracle AI Database and LangChain.
  2. My NEW FAVORITE Skill - Claude Code Drives My Whole Computer (Better Computer Use)Introduces a new 'Drive Screen' skill for AI agents that enables direct computer UI control. This skill allows LLMs to perform tasks such as morning setup, staging demos, and testing desktop applications without relying on external tools like Codex or specific harnesses. The skill is presented as lightweight, customizable, and easily integrated into existing projects or provided to coding agents. The post also discusses the underlying principles of LLM-driven computer interaction, the role of custom CLI scripts, and potential security concerns like prompt injection in this new paradigm.
  3. GPT-6 Astra Just Made AI Software Factories Real (Here's How to Run One)This post introduces and demonstrates the implementation of an 'AI Software Factory' powered by GPT-6 Astra. It details how to set up a remote server (VPS) using a single prompt, configure it with necessary tools (GitHub, Codex login), and then feed it PRDs or issues to generate fully validated and shipped code. The factory's workflow involves the LLM orchestrating tasks, calling external tools to manage rate limits and specific operations. The post provides links to the open-source factory repository, a server cheat sheet, and a related course on agentic coding.
  4. No One Talks Enough About Security for AI Coding. Here's How I Do It in My WorkflowsThis post introduces a security-focused workflow for AI-generated code. It highlights the vulnerability of AI-generated code and proposes a solution using Archon (an open-source harness builder) and SonarQube (a static analysis tool) as a mandatory gate within the AI coding workflow. The workflow ensures that security scans are performed and that code is only accepted if it passes these checks, addressing the limitations of relying solely on prompts or additional review agents for security. The contribution is the practical implementation of this secure AI coding workflow, demonstrating its end-to-end execution and the types of vulnerabilities it can catch.
  5. The Safest Way I've Found to Let AI Coding Agents Run Real InfrastructureThis post introduces a secure method for AI coding agents to interact with real infrastructure by leveraging Kestra's flow-based execution. Instead of granting direct credentials, agents are given a list of allowed flows defined in YAML and stored in Git. Each execution is logged as a readable event, significantly reducing the risk of unauthorized actions. This approach enhances the safety of AI agents operating on production systems.
  7. AI Software Factories Are the Next Big Thing (And I'm Building You One)This post introduces the concept and an open-source implementation of an 'AI Software Factory'. This factory aims to automate the entire software development lifecycle, taking a Product Requirements Document (PRD) as input and producing shipped code as output, with minimal human intervention in the code review process. The author has experimented with this 'dark factory' approach to build and maintain an AI tutor called Dynachat. The post highlights the underlying 'Archon' workflow engine and a 'skills' framework that powers the factory. An open-source version of the AI Software Factory is being made available, with the prompt for installation provided in the README. The author emphasizes this as a priority for future development and invites collaboration.
  8. 11 Tiny Coding Agent Fixes With A Stupid Amount Of PayoffThis post provides 11 practical tips for improving the reliability of coding agents. Key technical contributions include: 1. Emphasizing prompt engineering for agents rather than humans. 2. Highlighting the degradation of instruction files over time and the need for maintenance. 3. Advising against the use of `/compact` operations due to potential loss of safety constraints. 4. Recommending the placement of load-bearing rules within hooks for better reliability. 5. Suggesting that smaller context windows can be more effective than larger ones. 6. Explaining common reasons for hitting rate limits quickly. 7. Advising against escalating tasks mid-execution. 8. Debunking the need for explicit agent coordinators. 9. Stating that the writer should not approve their own work. 10. Warning against over-revision of agent output. 11. Framing validation as a continuous system rather than a single step.
  9. The Hidden Flaw of EVERY Coding Agent Now Has a SolutionThis post identifies a critical flaw in current AI coding agents: they are primarily evaluated on code correctness, not on adherence to business logic and access control. It introduces Sonar's Hunter Agent as a solution, which uses playbooks to understand application requirements and verifies AI-generated code against these rules, reporting violations as standard issues in SonarQube Cloud. This shifts the focus from 'is the code wrong?' to 'does the code do what it's supposed to do, including business rules and access control?'.