curated-claude-code

A small, vetted, self-evolving harness for Claude Code — curated, not dumped.

Most "everything for Claude Code" collections optimize for count — hundreds of agents, skills, rules, and event hooks that auto-run on your tool calls. It looks impressive and fails quietly: bloated context, overlapping triggers fighting to activate, and hooks that mutate things without asking.

This repo optimizes for the opposite. A handful of skills that earn their place, a spine of safety guards, and — the actual product — a system that improves itself over time, safely.

A harness that grows is worth more than a harness that's big. The thing that compounds isn't the block count — it's the intake gate, the hygiene loop, and the learning machinery. → docs/EVOLUTION.md is the heart of this repo; start there.

What this is / isn't

• ✅ A curated set of executable skills + always-on principles you can drop into ~/.claude/.
• ✅ A method (/vet) for deciding what to adopt, and a registry so you never evaluate the same tool twice.
• ✅ A self-evolution loop — corrections become memory, findings become skills, the stack stays clean.
• ✅ Safety by design — no surprise autonomous actions.
• ❌ Not a 500-block mega-bundle. Not a framework you fork. No auto-running hooks.

Looking for a "self-learning" setup?

Single-skill self-learning harvesters are trending, and the good ones are genuinely good. This repo is the system that pattern lives in: four learning layers (corrections → memory → skills → gated intake) instead of one harvest step, a promotion gate so unverified guesses never become authoritative skills, an intake gate (/vet) so the stack doesn't bloat as it learns, and a hygiene loop (system-health) so it doesn't rot. If you only want the harvest step, Kulaxyz/self-learning-skills does it well — we credit and cherry-pick it in docs/EVOLUTION.md. If you want learning as an operating discipline, start here.

Why curated, not dumped

• Vet before adopt. Nothing enters on hype. Every incoming tool runs through one pipeline — relevance → security as a blocking gate → duplication → real ROI → verdict. The security phase reads a skill's body as a behavioral instruction and checks the agentic risks (hidden instructions, prompt injection, trigger abuse, excessive agency) a code-only scan misses.
• It evolves. A user correction is the strongest learning signal; a finding is captured at discovery; a 5+-iteration procedure becomes a skill — but only past a promotion gate (verified check + named failure pattern + a ruled-out dead-end), so confident guesses never get enshrined as truth. The system compounds instead of going stale. → docs/EVOLUTION.md
• Cherry-pick over giant. A good idea inside a huge framework doesn't justify installing the framework. Extract the technique, leave the bloat.
• No auto-hooks. Auto-mutation without confirmation is exactly the failure mode to avoid. → rules/safety-guards.md

Full rationale: docs/METHODOLOGY.md.

What's inside

Skills (skills/)

Agents (agents/)

Two by design, not sixty — each reflects the verify-before-trust stance, not filler.

Rules (rules/)

Docs & examples

• docs/EVOLUTION.md — how the harness improves itself (start here).
• docs/METHODOLOGY.md — why curated, not dumped.
• docs/MEMORY_SYSTEM.md — a file-based, self-learning memory pattern.
• docs/TOOL_ROUTING.md — task → tool routing, proactively.
• docs/CHANGE_DISCIPLINE.md — backup → surgical → verify → record.
• examples/ — a registry template and a worked /vet verdict.

Which skill should I use?

Quick start

Install manually — copy only what you want. (No curl | bash: this harness practices what it preaches.)

git clone https://github.com/Sanexxxx777/curated-claude-code
cd curated-claude-code

# install the skills you want
cp -r skills/vet ~/.claude/skills/
cp -r skills/ship-secure ~/.claude/skills/
# …repeat for the rest

# agents → ~/.claude/agents/ ; rules → append into your CLAUDE.md or ~/.claude/rules/
cp agents/*.md ~/.claude/agents/

Restart Claude Code, then invoke a skill by its name (e.g. /vet).

Attribution

This harness was built the way it tells you to build — by cherry-picking techniques and crediting them:

• konsilium adapts Andrej Karpathy's llm-council (independent opinions → anonymized peer review → chairman synthesis).
• ship-secure adapts Prajwal Tomar's "Vibe Coders Are Getting Sued" launch playbook.
• code-principles is derived from Karpathy's widely-shared CLAUDE.md principles.
• goal's design phase cherry-picks loop-design ideas from ksimback/looper.
• teach adapts Anthropic's guided-learning prompt.
• The skill promotion gate in EVOLUTION.md cherry-picks the verification bar from Kulaxyz/self-learning-skills.

Work with me

I build and harden Claude Code workflows, agentic setups, and web apps — and I do managed hosting and site builds. If these patterns are useful to your team, or you want help applying them, reach out:

• GitHub: open an issue, or @Sanexxxx777
• Email: [email protected]

Contributing

PRs welcome — see CONTRIBUTING.md. The bar: it must pass its own /vet, ship no auto-hooks, contain no secrets, and stay surgical. Curation is the point; volume isn't.

If this saved you time, a ⭐ helps others find it.

License

MIT © Aleksandr Shulgin (@Aleksandr_NFA) · @Sanexxxx777