My Claude setup
One terminal, set up as a workshop. These are the skills, subagents, hooks and tooling I run inside Claude Code — the stuff that turns 'chat with an AI' into a system that actually ships. Most are public skills and plugins you can install today; the few I wrote myself go up as I open the repos. Steal anything.
Skills
Reusable capabilities I trigger by name. Some I built, some are from the community.
caveman
Compresses responses to roughly a quarter of the tokens by speaking terse — with lite / full / ultra levels — while keeping every bit of technical substance.
session-handoff
Captures a long session into a handoff doc and rolls into a fresh one before context auto-compacts, so nothing gets silently truncated.
cost-ack
Stops and makes me acknowledge cost before anything bills outside my flat plan. A guardrail against surprise spend.
save-to-spotify
Turns a script into a narrated audio episode and publishes it — TTS, chapters, cover art.
verify
Actually runs the app and observes behaviour before claiming a change works. Evidence over assertion.
loop / schedule
Run a prompt on an interval, or hand a recurring task to a scheduled agent. Automation for the automation.
Subagents
Specialized agents I delegate to, so the main thread stays focused and my context stays clean.
Explore
A fast read-only scout for 'where is X / which files touch Y' — it reads the codebase so the main thread doesn't have to burn context.
code-reviewer
Reviews fresh changes for correctness, convention and regressions.
code-simplifier
Untangles complex code while preserving behaviour.
ux-reviewer
Critiques UI, interaction and accessibility on anything user-facing.
Plan
Architects an implementation plan before a line of code is touched.
Hooks
Shell hooks that fire on events — the harness runs these, not the model, so behaviour stays consistent across every session.
SessionStart
Flips on caveman mode and restores any handoff session on resume.
UserPromptSubmit
Keeps caveman active and watches the context budget, nudging a handoff before auto-compaction hits.
PreToolUse (Bash)
A cost-guard that blocks actions which would bill outside my plan.
Plugins
Bundled extensions that add whole workflows.
superpowers
A big toolkit of dev workflows: brainstorming, test-driven development, systematic debugging, parallel-agent dispatch, plan execution.
frontend-design
Generates distinctive, production-grade UI that dodges generic AI aesthetics.
context7
Pulls current library and framework docs on demand, so answers track the latest API instead of stale training data.
code-review / code-simplifier
Pull-request review and code simplification, on tap.
skill-creator / claude-md-management
Author new skills; audit and improve CLAUDE.md project memory.
LSP (typescript / pyright / swift)
Language servers wired in for real type-aware navigation and checks.
MCP servers
External tools Claude can call directly through the Model Context Protocol.
context7
Live documentation retrieval for any library or SDK.
claude-in-chrome
Drives a real Chrome tab — navigate, click, read the page, screenshot, even record GIFs of a flow.
Curious how this site is built? The write-ups are all here, and the code goes up as I open the repos — the system underneath it all is the Operator Loop.