Agentic Coding Repo

A personal repo for agentic coding workflows — configurations, scripts, and skills for working with AI coding agents (primarily GitHub Copilot CLI).

What's in here

AGENTS.md

Instructions injected into every AI agent session. Enforces a git worktree workflow so multiple parallel agent sessions don't conflict with each other. Each session gets its own branch and worktree directory.

notes/

A collection of obsidian notes I've made during AI coding research.

scripts/

Shell scripts for managing git worktrees:

• createWorktree.sh — creates a new worktree + branch for an agent session
• removeWorktree.sh — removes a worktree and deletes its branch
• list_worktrees.sh — lists all active worktrees
• openWorktreeInCode.sh — opens a worktree in VS Code
• checkIfAlreadyInWorktree.sh — guard script to prevent nested worktrees

skills/

Reusable agent skills. Skills are invoked by name during a session to activate specialised behaviour.

Skill	Description
answer-and-stop	Answer the user's question briefly and directly, then stop the conversation
create-gh-issue	Create GitHub issues with the gh CLI. Assumes the existance of issue templates.
create-gh-pr	Create GitHub pull requests with the gh CLI. Assumes the existance of a pull request template.
grow-docs	Meant to be used at the end of coding sessions to document the codebase and changes made. It creates clusters of brief, wikilinked notes, then audits the vault for broken links and doc gaps to fill next.
handoff-to-worktree	Crystallise a conversation into a plan and hand it off to a fresh session
implement-strategy	Guides an implementation of a strategy. Mentions the docs and testable-module strategies as part of the implementation steps, and grills the user for any missing details to create a complete plan.
review-changes	A WIP code reviewer for multishot AI workflows.
testable-module	Inspired by the improve-codebase-architecture, it tries to refactor already existing code not the improve it, but to rewrite it in a more testable way. Using Test-Driven Development approach, it writes tests first and then refactors the code to make it pass the tests. Great for making the codebase more AI-friendly and improving test coverage.
caveman	Ultra-compressed communication mode (~75% fewer tokens)
grill-me	Relentlessly interview the user about a plan or design
domain-model	Stress-test a plan against the project's domain model and update docs
improve-codebase-architecture	Find deepening/refactoring opportunities informed by the domain language
write-a-skill	Write a new skill based on user instructions

The caveman, grill-me, domain-model, improve-codebase-architecture, and write-a-skill skills were copied from mattpocock/skills.

Potential additions

• Triage skill - would require a set of defined github label.
• More multishot components - workflows/skills (from triage to issue review, tests, implementation, docs update to the final PR, with intermediate code reviews and checks, adhering to the coding guidelines like module testability).
• More strict and interconnected core guidelines (e.g. "always write tests first (red/green)", "always update docs at the end", "read the related docs from...", "make sure the module is testable") and skills to enforce them.
• More defined worktree lifecycle management (e.g. automatic cleanup of old worktrees, reminders to delete worktree after session ends, etc.). Alternatively, use something else than git worktrees? Docker containers maybe?
• Docs expansions - something like a glossary for the calcified terminology and concepts - no more than one or two sentences per term, with links to examples and more detailed explanations in seperate notes. It's annoying that you have to reexplain the same concepts over and over again in every session, and a shared glossary would speed things up and improve consistency across sessions. Might want to combine it with skills like grill me and grow docs, or even mention the glossary and add an auto-updating instruction directly to AGENTS.md.