Claude Code skills

A small set of Claude Code skills I use day-to-day. Each one lives in its own folder under skills/ and is self-contained. You can copy any single skill into your own ~/.claude/skills/ without taking the rest.

What's in here

Building and shipping

• diagnose runs a disciplined bug-diagnosis loop: reproduce, minimise, hypothesise, instrument, fix, regression-test.
• improve picks one worthwhile change for the current project, proposes it, then either grills you on it or implements it.
• improve-codebase-architecture finds deepening opportunities and walks the design tree with you.
• karpathy-guidelines is a short set of coding behavioral rules to reduce common LLM mistakes, adapted from Andrej Karpathy's principles.
• frontend-design biases UI work toward distinctive, polished output. Reads your stack preferences from a local PREFERENCES.md.
• deploy ships the current project over SSH to hosts you list in HOSTS.md.

Planning

• grill-me interviews you about a plan or design until every branch of the decision tree is resolved.
• to-prd turns the current conversation into a PRD on your issue tracker.
• to-issues breaks a plan into independently-grabbable tickets using tracer-bullet slices.

Research and notes

• research-here does multi-angle research with URL-per-claim citations and writes the synthesis into ./research/ in the current repo.
• note captures content into an Obsidian vault. The vault path lives in a gitignored VAULT.md per machine.
• query asks a question against the same vault via a sub-agent, so the main session never loads vault context.

Workflow

• git-cleanup prunes merged local branches.
• caveman switches into an ultra-compressed communication mode (≈75% fewer tokens).

Using the whole set

git clone https://github.com/AbstractNucleus/claude-config.git ~/.claude
claude  # authenticates on first run

Skills that need a path or preference (deploy, note, query, frontend-design) will ask you for it the first time they run, and save your answer to a gitignored file in that skill's folder.

Using a single skill

If you only want one:

cp -r path/to/this/repo/skills/diagnose ~/.claude/skills/

Restart Claude Code. The skill self-registers via its SKILL.md frontmatter.

Repo layout

This repo doubles as a synced ~/.claude/ — cloning it there carries skills and config across machines. Tracked: skills/, CLAUDE.md, README.md. Everything else is gitignored: credentials, install state, plugins, conversation history, logs, caches, and settings.json. .gitignore is the authoritative list.

settings.json is per-machine, not synced — Claude Code rewrites UI/runtime toggles (editor mode, remote-control, notifications) back into it on every change, which caused constant cross-machine churn. So each machine keeps its own, and shared config (permissions, hooks, model) is set up per machine.

Before committing, scan git status for secrets or unexpected new files (.gitignore can lag new Claude Code additions). If a secret ever lands in a commit, rotate it before rewriting history.

Design notes

• Standalone. No skill calls another skill by name. You can delete or copy any folder without breaking the rest.
• No machine-specific values committed. Per-machine config (vault paths, deploy hosts, design preferences) lives in gitignored files alongside each skill, with a committed .example.md template.
• No plugin dependencies. Everything you need is in this repo. settings.json doesn't enable any plugins.

Credits

karpathy-guidelines is adapted from forrestchang/andrej-karpathy-skills (MIT).

License

MIT. See LICENSE.