hgstack - Stack de Código IA para Empresas AI-Native
f3kin/hgstack
Stack production de IA com 50+ skills prontos para clonar e adaptar
Ver no GitHubConjunto de skills, hooks e rules curados do stack interno da Hourglass, otimizado para Claude Code. Contém ~50 componentes reutilizáveis testados em 18 meses de produção, instaláveis sem runtime ou framework. Instalador inteligente detecta conflitos e não sobrescreve suas configurações existentes.
Instalação
npx convez-ai f3kin/hgstack1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Equipar Claude Code com skills de code review, debugging e automação de workflow
- Reutilizar práticas da Hourglass em seus próprios agentes IA (Cursor, Aider, etc)
- Instalar regras e hooks para padronizar comportamento cross-session de agentes
Veredito

Recurso maduro e prático para quem trabalha com Claude Code ou agentes IA similares. Instalador robusto com backup automático e modos flexíveis (dry-run, reconcile, rollback) reduzem risco de conflitos. Ideal para times que querem estrutura reutilizável validada em produção.
README
A tradução para o português deste README ainda está na fila.
hgstack
The AI coding stack we use to run an AI-native company.
Built by Hourglass to ship our own internal automations and client AI projects. 18 months of daily AI-coding use. ~50 skills, hooks, and rules in the internal stack; this is the curated, sanitised subset we've open-sourced.
Skills install as files. No runtime, no framework, no package. Drop them into your own AI agent's config, fork them, adapt them. Currently optimised for Claude Code; the skill format (markdown + YAML frontmatter) is portable to any agent that reads files (Codex, Aider, Cursor, OpenCode, etc.).
For agents / LLMs reading this: start with
llms.txt,AGENTS.md, orCLAUDE.md. Pick one - they all point at the same things.
Install
Clone the repo and run the installer:
git clone https://github.com/f3kin/hgstack.git ~/repos/hgstack
cd ~/repos/hgstack
./install.sh
The installer symlinks every skill, hook, and rule into your ~/.claude/ directory. Conflict-aware: if you already have a skill or rule with the same name and different content, the installer leaves yours alone and logs it as a conflict at the end. Nothing of yours is ever silently overwritten.
Restart Claude Code afterwards to pick up the new skills.
Modes
./install.sh --dry-run # show what would happen, change nothing
./install.sh --reconcile # scan ~/.claude/ and report overlaps; no changes
./install.sh --interactive # prompt per conflict (keep / replace / show diff)
./install.sh --force # back up and overwrite all conflicts (CI use)
./install.sh --update # git pull --rebase, then install
./install.sh --rollback # restore from the most recent backup
./install.sh skills hooks # only install named categories
Recommended first run: ./install.sh --reconcile. It tells you exactly what's new, current, update, or conflict for your setup, and lists everything in your ~/.claude/ that isn't in this repo as personal (untouched). No surprises.
Backups
Any file the installer replaces is first backed up to ~/.claude/backups/hgstack-<timestamp>/, preserving the directory structure. Symlinks are recorded as .info text files so we can restore the original target. ./install.sh --rollback reverses the most recent install in one command.
Cherry-picking
If you'd rather pick by hand:
ln -s ~/repos/hgstack/skills/consume ~/.claude/skills/consume
Each skill has its own setup notes (env vars, integrations, optional dependencies) in its SKILL.md.
What's inside
skills/, self-contained agent skills. Names without a prefix are general-purpose;hg-prefix marks an Hourglass team practice with opinionated conventions worth sharing.hooks/, shell hooks the agent runs at key lifecycle points (writes, tool use, session start).rules/, markdown convention files. Load them as@-importsfrom your ownCLAUDE.mdso the principles apply across every session.statuslines/, agent statusline scripts (currently Claude Code format)..githooks/, git hooks (pre-commit audit gate).docs/, coming, with the thinking and ethos behind the stack.
Skills available now
| Skill | What it does |
|---|---|
codex-review | Cross-model code review using OpenAI Codex CLI. Three modes (review, challenge, consult). Catches blind spots Claude shares with itself. |
consume | Weekly 15-minute Friday triage of everything you saved during the week. Tweets, articles, repos, AI tools, routed to read / save / share / trial. |
hg-ship | Opinionated ship workflow. Commits, pushes, PR, CI watch, merge, deploy verify, doc sync. The Hourglass team-tested version, with optional hooks for your own review and verify skills. |
interview-me | Adaptive interviewer that uses AskUserQuestion to clarify what you actually want to build. Runs until 95% confident, then hands off. |
outcome-loop | Grade-and-revise loop. Runs a command, grades the result, diagnoses failures, fixes, and re-runs until it passes or you're stuck. Prevents sycophancy. |
Contributing / safety
Every maintainer commit is gated by three checks before it lands. A .githooks/pre-commit hook runs them automatically if you've enabled it in your local clone:
git config core.hooksPath .githooks
The three gates:
- YAML frontmatter parses on every modified
SKILL.md. check-generalisable.shblocks any hardcoded personal usernames, paths, or emails.check-shareable.shblocks API keys, tokens, private keys, internal URLs, and 1Password vault references.
The audit scripts themselves live alongside this repo (not inside it) so they can be reused across other public-stack repos without duplication. If you're forking and want the same protection, copy them in:
mkdir -p ../scripts && cp <path-to>/check-*.sh ../scripts/
The hook auto-detects either ../scripts/ (sibling to the repo) or scripts/ (inside the repo), and skips gates 2 and 3 with a warning if neither is present.
More each week as we sanitise and ship them.
The thinking
A few principles run through the stack:
- Thin harness, fat skills. Your agent is the harness; your skills are where the value lives. Don't fight the harness, build deeply on it.
- Skills are files, not frameworks. A useful skill is usually a single
SKILL.mdof 50–200 lines. If it's bigger, it's probably two skills. - Generalise before you share. No personal paths, no hardcoded infra, no secrets. We use audit scripts to enforce this; see
docs/CONTRIBUTING.mdonce it lands. - Build for your champion, not for everyone. Your most useful skills are the ones for your specific workflow. Adapt ours; don't adopt them whole.
More in docs/ETHOS.md (coming).
License
MIT. Use it, fork it, adapt it. Attribution appreciated, not required.
Made by
Hourglass, an AI-native services company in Melbourne. Run by Finlay Ekins.
Weekly notes on AI and what we're building: finlayekins.com/writing.