Claude Kit
Live usableThe Claude Code setup I built this portfolio with: rules, agents, hooks and a status line that does not lie, the same on every machine.
My Claude Code setup, published and sanitised: lane-based routing rules, six subagents, ten skills, five hooks and a status line that tells the truth about context, cache and the HANDOFF. Mirror scripts keep it identical on every machine without deleting anything.
Technology
- AI & agents
- Automation
Sector
- Dev tools
Problem
I used Claude Code all day on three fronts and the same problems showed up in all of them: the most expensive model doing a small model's job, sessions losing the thread, and finding out too late that context was huge, the cache was cold or the handoff was stale.
Approach
One repo is the source of truth for what Claude Code reads on every machine: CLAUDE.md, rules,
agents, skills, hooks, the status line and rules/thresholds.json, the only place with numbers.
sync-to-home.sh copies it into ~/.claude with a backup, merges only the portable
settings.json keys and never deletes; verify.sh checks it landed and runs the three suites.
The bar reads the JSON Claude Code hands it on every turn and draws two lines: context in absolute
tokens, the cache state in one word (caliente, enfriando, fría) plus its minutes, plan
quota, git and how fresh the HANDOFF is. A missing field is painted ⚠, never 0. Lesson 01 of
the Code like me route walks through it sector by sector. The status line simulator (the demo link) shows
it reacting to controls: every state is the script’s real output, pre-rendered.
For anyone with more than one machine, the optional enjambre module (--with enjambre) adds a
hook that blocks a raw ssh to the nodes and makes every delegation go through enjambre’s queue.
Limits: the bar takes about 476 ms on Windows, over the 200 ms target; the kit’s CLAUDE.md is
its author’s own rules and is meant to be edited. Code under Apache-2.0.
Result
- automated tests passing
- 140
- statusline 67 + hooks 38 + enjambre module 23 + /route 12, Windows, 2026-10-15
- full status-line render on Windows + Git Bash
- 476ms
- docs/statusline-internals.md, 2026-09-22
- session lifecycle hooks
- 5
- home/.claude/hooks, 2026-10-15
- skills
- 10
- home/.claude/skills, 2026-10-15
- subagents by lane
- 6
- home/.claude/agents, 2026-10-15
What I'd do next
- Cut the bar's render time or accept the budget: 476 ms measured against a 200 ms target; the Mac is not measured yet.
- Port the ideas to Codex CLI and Gemini CLI.
- An UNATTENDED: line so a session can accept an approved scope without stopping to ask.
Architecture
Claude Code configuration kit: rules, agents, hooks and a status line mirrored to every machine without deleting anything.
- PC Windows
- Mac mini
- GitHub
Flows you can replay
- Instalar o actualizar el kit
- Una sesión lee las reglas y un hook actúa
- La barra de estado se repinta
- Módulo enjambre: la guardia de ssh
- Publicar la versión pública filtrada
How it was made
- 01 — Read your bar: what Claude Code is telling you on every turn — Look at the two lines of the status bar and know, at a glance, four things: how much context you have left before the session degrades, whether the cache is still warm, whether your work is safe (commit, push, HANDOFF)…
- Hooks — Five scripts under home/.claude/hooks/, wired in home/.claude/settings.json. Every hook always exits 0 — a hook must never be the reason a session, a compaction, or the tool it watches fails.
Sibling projects
- enjambre Agentic coding spread across hardware I own: a Windows PC orchestrates and a Mac mini runs Claude Code jobs headless, each in its own git worktree. Every job comes back as a reviewable branch and one line in an append-only ledger, and a local visor shows nodes, jobs and telemetry.
- Cert Studio A static practice app for the CCAR-F and CCDV-F certifications: mocks that spread questions by each domain's official weight, a timer, a scaled 100–1000 score and every wrong option tagged with the kind of trap it represents. Progress lives only in the browser. An unofficial study aid: all 240 questions are original and none reproduces a real exam item.