sskillsbaseGitHub ↗
↑ 22 skills · 8 agents · 2 hooks

Clone it. Claude’s ready.

One repo sets up Claude Code: skills, subagents, two light hooks and the global instructions. Plain files you can read.

Set up the itsjustiago Claude Code build: clone https://github.com/itsjustiago/skillsbase and follow setup/AGENT-INSTALL.md — check what I already have in ~/.claude, show me the options, and install the one I pick.

Paste into Claude — it asks how you want it

Project kickoff1Design4Engineering4Quality & security7Ship & sessions4Discovery2
The build

What one clone gives you.

Skills, agents and guardrails, all in one repo. Paste one prompt on a new machine and Claude is configured.

Global · always on

22 skills

Plain files in ~/.claude/skills, available the moment you open any project.

  • 6 written by us, vendored in the repo
  • 16 pulled from their authors' repos at install
  • 6 groups, from kickoff to quality & security
Subagents

8 agents

One writes the code; the others map, research, review, test and audit. financas covers Portuguese tax and money.

  • 2 have edit tools: engenheiro writes code, financas writes research notes
  • 6 have none: they map, research, review, test and audit
  • Prompts are written in Portuguese
Guardrails + instructions

4 pieces

Rules that apply in every session, from the hooks to the global CLAUDE.md.

  • hook: bloquear_kill_por_nome
  • hook: estado_sessao
  • instructions: setup/CLAUDE.md
  • settings: settings.json
Always on

Skills in every project.

Set up once by setup.sh, then available everywhere. 6 are ours; 16 are pulled from their original sources at install and never vendored.

Project kickoff

ui-ux-pro-max

external

Design direction for a new project: style, palette, type. Manual only, never auto-triggers.

nextlevelbuilder/ui-ux-pro-max-skill

/ui-ux-pro-max

Design

frontend-design

external

Aesthetic direction that avoids templated defaults. Fires on UI work.

anthropics/skills

impeccable

external

Design process: critique, audit, polish.

pbakaus/impeccable

/impeccable

emil-design-eng

external

Motion and the invisible details. Fires on UI work.

emilkowalski/skills

review-animations

external

Strict motion review. Default is to flag.

emilkowalski/skills

/review-animations

Engineering

systematic-debugging

external

Root cause before any fix.

obra/superpowers

debug X

verification-before-completion

external

Evidence before claiming done.

obra/superpowers

verify

supabase

external

Official Supabase guidance: auth, RLS, migrations, CLI.

supabase/agent-skills

supabase-postgres-best-practices

external

Postgres performance, the official way.

supabase/agent-skills

Quality & security

browser-testing-with-devtools

external

Real-browser testing via the Chrome DevTools MCP: DOM, console, network.

addyosmani/agent-skills

security-and-hardening

external

Hardening for input, auth, storage and integrations.

addyosmani/agent-skills

core-web-vitals

external

Fixes for LCP, INP and CLS.

addyosmani/web-quality-skills

web-accessibility

external

WCAG 2.2 audits and fixes.

addyosmani/web-quality-skills

semgrep

external

Static analysis scans with parallel subagents.

trailofbits/skills

differential-review

external

Security-focused review of PRs and diffs.

trailofbits/skills

supply-chain-risk-auditor

external

Flags dependencies at risk of takeover or exploitation.

trailofbits/skills

Ship & sessions

ship

own

commit, push, open PR, in one shot.

/ship

ship-merge

own

ship + CI wait + light review + squash-merge + cleanup.

/ship-merge

session-handoff

own

A clean end-of-session handoff before /clear.

wrap up session

preclear

own

Saves the task state and clears the chat. The next chat resumes from it.

/preclear

Discovery

skill-scout

own

Finds new skills, plugins and MCPs in the wider ecosystem.

watch-youtube

own

Transcript plus key frames from YouTube videos.

Orchestrator model

One briefs. One builds.

The main session briefs and reviews. The engenheiro writes the code; the other agents map, research, review, test and audit it.

  1. 1

    Main session

    Talks to you, briefs the work and reviews the diff.

  2. 2

    engenheiro

    Writes the code from the brief.

  3. 3

    Other agents

    Map, research, review, test and audit. design and seguranca are strictly read-only.

The 8 subagents

engenheiro

Builds

Writes the code from a complete brief: features, fixes, UI, bulk edits. Prefers a worktree when it touches many files.

explorador

Maps

Finds where things live and returns file:line before anything is changed.

investigador

Research

Researches tools, repos and libraries; verifies claims and ranks options with sources.

revisor

Review

Reviews code before merge and points at concrete problems with file:line. No praise, no rewrites.

testador

QA

Runs the app for real: navigates, clicks, reads console and network, takes screenshots.

design

Audit · read-only

UI critique: hierarchy, tokens, type, spacing, motion, accessibility. Proposes fixes, never applies them.

seguranca

Audit · read-only

Security audit: exposed secrets, injection, authz, RLS, dependencies, public surfaces.

financas

Money research

Portuguese tax and investing research from official sources, with auditable maths. No personal advice, no orders.

Guardrails & instructions

  • bloquear_kill_por_nome

    hook

    PreToolUse hook on Bash: blocks killing processes by name. Needs python3.

  • estado_sessao

    hook

    SessionStart hook: injects the folder's state file on startup, compact and clear, so work resumes without re-explaining. Needs python3.

  • setup/CLAUDE.md

    instructions

    The global ~/.claude/CLAUDE.md: language, git and ship rules, orchestrator mode.

  • settings.json

    settings

    Created only if missing. Registers the hooks, enables 1 plugin (security-guidance), auto-compacts at 250k and denies the Artifact tools.

How it's made

Plain files, pulled from source.

Rebuilt from scratch as files. The per-project catalog, the install profiles and the heavy MCP set are gone: everything left has to earn its start-up cost.

22
global skills
8
agents
16
pulled from source
2
hooks

Files, not plugins

Skills, agents, the hooks and the instructions are plain files in ~/.claude. It works on the desktop app without the claude CLI on PATH, and updating means re-running a script. The only plugin is security-guidance, enabled by the bundled settings.json when you have none.

Third parties from source

The 16 external skills are never vendored. install-externals.sh clones each upstream at install time, so authors and licenses stay theirs and updates are a re-run.

Orchestrator model

One main session briefs and reviews. The engenheiro agent writes the code; the read-only agents map, review, test and audit it.

Safe by default

Two light hooks: one blocks killing processes by name, one restores the session state on start. No per-edit hooks to add latency. Setup is additive, sync is a dry-run until --apply, every replace leaves a backup and settings.json is never overwritten.

The full reasoning, what was cut and why, lives in DECISIONS.md; the install logic is setup.sh, and the overview is the README.md.

Install

One line. Claude does the rest.

Paste this into Claude. It checks what you already have, asks how you want it, and installs — nothing you didn't pick.

Set up the itsjustiago Claude Code build: clone https://github.com/itsjustiago/skillsbase and follow setup/AGENT-INSTALL.md — check what I already have in ~/.claude, show me the options, and install the one I pick.

Then Claude asks you how:

Skills only, on top

recommended

Installs the 22 skills, 8 agents, and the hook files. Keeps your other skills, agents, plugins, CLAUDE.md and settings. Same-name items are replaced, with a backup in ~/.claude/backups/.

Skills + instructions

Full bootstrap: everything above, and your CLAUDE.md is replaced (backup kept). Creates settings.json only if you have none.

Match the build

Installs everything, then removes the global skills, agents and hooks that aren't in the build. Also replaces your CLAUDE.md (backup kept). Never removes plugins or settings. Dry-run first, backup on apply.

Instructions only

Just the global CLAUDE.md, with a backup of your current one. Skills, agents and hooks untouched.

  • Needs git and node, plus python3 for the hooks. Restart Claude Code when it finishes.
  • The instructions are the owner's: auto-merge and push without asking, written for "Tiago". Adapt them after installing.
  • settings.json is never overwritten. Registering the hooks is proposed, and only done with your OK.
  • External skills are cloned from the upstream HEAD, unpinned by design. Review setup/install-externals.sh first.

Fresh machine or a setup of your own, it adapts. Everything comes from one public repo: github.com/itsjustiago/skillsbase.