Writing and Reviewing Agent Skills - Common Pitfalls
The top pitfalls I keep seeing when reviewing Agent Skills, and what to do instead
smcleod.net
Recent content on smcleod.net
Latest posts
Measuring Model Quantisation Quality with KL Divergence
Using KL divergence to rank MLX quantisations of Qwen 3.6 (27B dense and 35B-A3B MoE)
New Apple Silicon M4 & M5 HiDPI Limitation on 4K External Displays
A regression in external display support on Apple Silicon M4/M5 generation chips prevents full HiDPI modes on 4K monitors, limiting users to either blurry non-HiDPI or reduced workspace HiDPI.
The advice I find myself repeating every time someone asks how to get started with Claude Code
I spend a lot of my time helping people who are getting started with Claude Code. These are the key things I find myself repeating.
Patching NVIDIA's driver and vLLM to enable P2P on consumer GPUs
NVIDIA artificially restricts peer-to-peer (P2P) GPU communication to their enterprise cards. Turns out this is a software limitation, not a hardware one. I patched my drivers to remove it, hacked vLLM to take advantage of it, and got a 15-50% throughput improvement running Qwen 3.5 35b on dual RTX 3090s.
The Role Bridging Problem
An observation on functional correctness without domain quality.
Stop Polluting Context - Let Users Disable Individual MCP Tools
If you’re building MCP servers, you should be adding the ability to disable individual tools.
MCP DevTools
A single, modular MCP server for AI coding agents.
Agentic Coding Adoption Cost Cycle
The two common themes I see with engineers adopting agentic coding tools
Agentic Coding Workflow & Cline Demo
Recording of a demo and Q&A session on my Agentic Coding workflow and Cline use.
Vibe Coding vs Agentic Coding
From Creative Exploration to Production Quality
My Plan, Document, Act, Review flow for Agentic Software Development
An overview of my Setup, Plan, Act, Review & Iterate workflow for agentic software development.
LLM Sampling Parameters Guide
A practical guide to LLM sampling parameters for Ollama, llama.cpp, and MLX
Getting Started with Agentic Systems - Developer Learning Paths
A curated learning path for engineers looking to gain practical experience with AI and agentic systems.
The Cost of Agentic Coding
The cost of not leveraging agentic coding is likely far greater than you think.
The Democratisation Paradox: What History Teaches Us About AI
The Democratisation Paradox: What History Teaches Us About AI.
The effects of prompt caching on Agentic coding
What is prompt caching and why is it so important for Agentic coding?
Agentic Coding - Live Demo / Brownbag
Video recording of a brownbag presentation / live demo I ran on Agentic Coding using Cline/Roo Code
Bringing K/V Context Quantisation to Ollama
K/V context cache quantisation has been added to Ollama. This enables significant reductions in VRAM usage, allowing users to realise the potential of expanded context sizes and run larger models at their existing context sizes.
Will AI Take My Job?
It's probably not so much AI itself you have to fear.
Generating Diagrams with with AI / LLMs
Generating diagrams with AI / LLMs
Ingest: Streamlining Content Preparation for LLMs
A CLI tool for parsing directories into LLM-friendly markdown
LLM Parameter Playground
A web app to explore the effects of different inference parameters on LLMs.
Code, Chaos, and Copilots (AI/LLM Talk July 2024)
Integrating AI into Daily Development
Understanding AI/LLM Quantisation Through Interactive Visualisations
AI/LLM Quantisation Visualised
Rating AI Tools
Subjective ratings of the notable AI tools I've tried.
Gollama: Ollama Model Manager
Gollama is a Go-based client for Ollama for managing models.
Confuddlement: Download Confluence Spaces as Markdown, Summarise with Ollama
Download Confluence Spaces as Markdown, Summarise with Ollama
NVApi - Nvidia GPU Monitoring API
Nvidia GPU Monitoring API
SuperPrompter - Supercharge your text prompts for AI/LLM image generation
SuperPrompter is a Python-based application that utilises the SuperPrompt-v1 model to generate optimised text prompts for AI/LLM image generation (for use with Stable Diffusion etc…) from user prompts. See Brian Fitzgerald’s Blog for a detailed explanation of the SuperPrompt-v1 model and its capabilities / limitations. Features Utilises the SuperPrompt-v1 model for text generation. A…
Llamalink - Ollama to LM Studio LLM Model Linker
This is a simple command-line tool that allows you to easily link Ollama models to LM Studio's directory structure. It simplifies the process of symlinking Ollama models to LM Studio, making it convenient to use the models in both applications.
Open source, locally hosted AI powered Siri replacement
Using an ESP32 S3 Box 3 with Willow, Home Assistant and Large Language Models (LLMs) create a locally hosted, offline, AI powered Siri / Alexa replacement.
Fun with Makefiles - Dynamic Menu Generation
This reusable snippet will generate a menu of targets from the Makefile. It will use `fzf` if it is installed, otherwise it will use a numbered menu.
MBA Washing
"MBA Washing" refers to the phenomenon where individuals or companies, often with a strong theoretical or academic background but limited recent practical experience, adopt and reinterpret industry-specific terminology and cultural movements.
SDXL LoRA Training
A talk I gave to some peers on creating your own SDXL LoRA models from my tinkering around over the last few weeks.
Fixing AMD CPU Scaling on Fedora
Enabling the new AMD P-State Driver on Fedora with Kernel 6.4
Defaulting git clone to shallow (depth=1)
Before adding this to my shell config, I would manually add –depth=1 to all my git clones.
How likely would you be to block a company from asking you to rate everything they do?
NPS Surveys are corporate spam
Escaping the Time Trap: Why Estimating Effort, Not Time, Leads to Greater Success
Effort is non-linear and multi-dimensional. Time (with relation to delivery) is linear and one-dimensional
Introduction to AI and Large Language Models (LLMs)
A high level intro to LLMs that I'm writing for a few friends that are new to the concept. It is far from complete, definitely contains some errors and is a work in progress.
Open Budgeting - Subscriptions & Licences
Publicly documenting my subscription and license expenses
Energy, Sustainability and Deployment Frequency
Deployment / Delivery Frequency I often end up needing to advocate for more frequent delivery/deployments with clients. There’s the usual benefits commonly discussed such as improved feedback, reduced risk, well understood processes, maintainable dependencies etc… however what’s often missed entirely is how it relates to the health and sustainability of the team.
Zsh Configuration and Plugins - Part Two
My Z-Shell configuration, scripts and hacks
Github Not-So-Reusable Actions
Github Actions Reusable Workflows vs Composite Actions
YAML Anchors and Aliases
Using YAML Anchors and Aliases to make config files more DRY
Zsh Configuration and Plugins - Part One
My Z-Shell configuration, scripts and hacks
Encourage Your Peers To Contribute To Open Source
Contributing to Open Source is important to the quality and maintainability of the software and engineering communities we rely on every day - so why is it that so many developers/engineers never participate?
Making Work Visible - Avoid DMs
We create more value by having conversations in public instead of behind closed doors.
The Best Of - 2022 Edition
Near the end of each year I note down a summary of the best apps I’ve enjoyed using throughout the year, here’s 2022.
It's 2022 and we're (still) not deploying enough
We’re (still) not deploying enough It’s 2022 and not deploying frequently enough is still one of the most common causes of software failure.