sammcj / agentic-coding

Sam's Agentic Coding Rules, Templates and Examples • Sam's Agentic Coding Rules, Templates and Examples • Coding Agent Rules, Skills and Templates • Patterns / Workflows • MCP Servers (Agent Tools) • Tips For Agentic Coding • Links • License A collection of coding rules, templates, MCP servers and examples for working with Agentic Coding tools Coding Agent Rules, Skills and Templates The repository is organised with tool-agnostic content at the root level: • Agent Rules - Agent instruction files • Rules/CLAUDE.md is usually the most up to date and comprehensive • Rules/AGENTS.md is more generalised for use with other agentic coding tools. • Agent Skills - Reusable skills that extend agent capabilities • Agent Commands - (Prompt templates) • Hooks - Hooks that can be used when specific lifecycle events occur (e.g. after writing to a file) • Custom Agents - Custom defined agents. • Claude Code Statusline - Custom statusline for Claude Code that shows usage and context information. • Claude Local Marketplace - My local marketplace (plugins) for Claude Code, including LSP definitions. Note: Most rules and skills are portable between agentic coding tools. The exception being Claude _Skills_, which I'm yet to see other tools support - but have no doubt they will as they're one of the most powerful and useful agentic coding concepts. Agentic Coding Tools - Not All Created Equal **I _highly_ recommend Claude Code as the best agentic coding tools.** In my experience Claude Code _far_ outperforms the many other tools I've worked with over the past 3+ years including: Copilot Agent, Gemini CLI, Codex, Cursor, Kiro, Windsurf, Augment Code, AntiGravity, Aider and many, many others. The next best thing to Claude Code? I think that's probably OpenCode or Cline. • Claude Code offers the fastest agentic coding experience that leads to the highest quality outputs, it's often a trend setter when it comes to new features and provides the best value for money if you can pair it with the Claude Max 5 or Max 20 subscription ($100-200 USD / month) which is far better value for money than what you would have to spend with consumption based pricing (~$2000-$3000USD+GST/mo). • OpenCode and Cline are the best options for using with self-hosted and alternative models, they're also Open Source. When setup correctly you can do fantastic work using Claude Code, Cline or OpenCode, but beware of other tools especially those such as Cursor or Kiro that have flashy marketing and offer low monthly subscription costs on paper - they tend to be severely limited in terms of their real capabilities and any tool offering sub $100USD/month plan is going to be heavily rate limited. I have a table where I'm trying to keep track of coding agent tools here: https://smcleod.net/agentic-coding-tools/ Patterns / Workflows I've shared (somewhat out of date) overview of workflow (Setup -> Plan -> Act -> Review & Iterate) in a blog post here (smcleod.net). The TLDR is that this is a lightweight approach of having coding agents help me design a documented plan with the goal, context, requirements / constraints and a phased checklist of tasks to complete and have them update the plan as they work through the tasks in each phase. Note: I've tried full blown '_Spec Driven Development_' but find it can be overkill for small-medium size projects where it can lead to over-engineering and significant documentation debt. MCP Servers (Agent Tools) For 99% of my work I only use a single MCP server: MCP DevTools I wrote it to provide the most common tooling I use with Agentic Coding. • Always have tools available to the agent that allow it to: • Search the web (if the agentic coding tool doesn't have this built in like Claude Code does). • Efficiently retrieve web page content as markdown (likewise here). • Lookup package documentation. • Perform correct math calculations. • Be mindful of how many tokens each MCP server adds to your context window ( in Claude Code), some tools abuse their descriptions and pollute the context window - for example Github's official MCP server alone uses ¼ of the entire context window of Claude Sonnet 4.5. • If the tool or knowledge you want to give an agent access to already as an easy to use CLI tool I'd recommend simply instructing your agent to use the CLI tool when it needs to rather than providing a full blown MCP server for it. --- Tips For Agentic Coding Claude Code • Use plan mode to plan tasks for Claude to work from before diving into making changes. • Lean into skills, dynamical context acquisition is very powerful. • Encourage Claude to use sub-agents where it makes sense to do so. • Create a command (prompt template) that prompts Claude to perform a critical self review and to fix any issues it may find, use this after it completes work. • Adding a rule to remind Claude to pipe the output of potentially noisy shell commands to null can help reduce token usage. • Adding your context usage, and limits consumption (if you're on a plan) to the statusline is useful. • Using voice to text with a tool like Handy to dictate work to Claude Code is super useful, sometimes with voice you capture intent that you'd otherwise edit out. • If Claude gets stuck on a complex issue, get it to stop and perform a systematic debug of the issue. • Setup your permissions (in settings.json) to pre-approve/deny/ask commands and file paths that Claude may want to use. If you're not sure of the syntax the docs are OK but you can also ask Claude for the correct way to write them. • Use Anthropic's official PPTX/DOCX skills to create documents. • Use the Excalidraw MCP to create diagrams. • Create shell aliases for the various claude CLI commands you use, e.g. , , etc. Writing Rules • **Only enable the rules you actually want to use** • Think about the signal to noise ratio of your rules (and context in general), _how much information could you be told at once and remember?_ • If you have a lot of rules that are not…