~/tags $ grep -l '#llm' _posts/*
// #llm
7 posts tagged.
Auditing a Homelab With AI Agents: Personas, Four Waves, and an Adversarial Verifier
The architecture of a security-review system built from AI agent personas: a read-only auditor with a hard constraint list, a four-wave pipeline where each wave answers a question the previous one cannot, a checklist ...
Relaxing an Over-Zealous AI Guard: A Silent Abort, a Three-Character Bypass, and an Empty Rule List
My Cline PreToolUse hook was ending whole conversations with no error message, because a tool_call hook that returns cancel aborts the entire agent run and discards the reason on the way, rather than denying the singl...
Lessons That Arrive On Time: A Scored Corpus, and What Six Phases Actually Cost
Continuing from "When the Output Looks Right": building discernment into a coding agent. A dated lesson in a rules file is useless at the moment it matters, so I moved them into a scored corpus that injects the releva...
When the Output Looks Right: Why a Green Test Suite Stops You Looking
Anthropic's AI Fluency Index found that people get less critical exactly when an AI produces something polished: fact-checking drops 3.7 points and spotting missing context drops 5.2 points the moment code or document...
From One Agent to a Team: Async Loops, Multi-Agent Personas, and an Autonomy Envelope for Cline
The next step after steering a single coding agent is letting it run unattended and splitting it into specialised personas. Here is the honest, safety-first version: a GREEN/AMBER/RED action envelope that defines what...
Teaching Cline to Learn From My Corrections: Turning Conversation Steering Into Durable Rules
Self-learning AI agents promise to get smarter from every interaction. Here is the honest version for a coding agent like Cline: it does not retrain its own weights from your chats, it consolidates your corrections in...
How I Steer Cline With a Tiered Rules System, a Persona, and a Pre-Commit Hook That Stops Me Leaking My Homelab
A 700-line steering layer for Cline split into always-on rules and on-demand reference docs, a banned-words list that kills the AI smell, opsec rules that keep internal hostnames out of public commits, a single canoni...