Skip to content

dimitris-di/lude-kit

Lude Kit

A library of senior grade Agent Skills for Claude Code and OpenAI Codex.

Drop in a whole engineering org. Dispatch the right specialist when you need them.

License Skills Spec Status


The pitch in 30 seconds

General purpose AI coding agents are competent at everything and excellent at nothing. They read a lot but never shipped a feature. Lude Kit fixes that. Every skill is one role or one job, written to the bar of a senior practitioner, with a precise trigger description that activates it only when relevant.

Install the library once, and the right specialist shows up automatically:

> "Help me design a billing system that handles refunds and chargebacks."
  → staff-software-architect + fintech-engineer activate.

> "This Postgres query is slow. EXPLAIN ANALYZE attached."
  → postgres-expert + senior-performance-engineer activate.

> "We had a sev2 yesterday. Need the postmortem by Friday."
  → postmortem-author + incident-commander activate.

> "Build a RAG pipeline over our docs with a real eval set."
  → senior-rag-engineer + senior-eval-engineer activate.

Compose them into multi agent flows and your orchestrator hands the work to the right person at every step. No prompt bloat, no generic answers, no role confusion.

Why this exists

A modern AI agent is a generalist who needs a senior on the team. Without one, you get answers that are technically correct but operationally naive: code that works in dev and burns in prod, designs that look good on a whiteboard and fall over at scale, security advice that's textbook and theatre. The fix is not a longer prompt. The fix is the right specialist on demand.

Lude Kit is that bench. Seventy senior practitioners, each focused on one role or one job, each with strong opinions about how that work is done well. Drop them into your environment and the work gets better immediately.

Status

Public alpha. Open sourced under Apache-2.0. Seventy of a planned one hundred skills landed; bar holds across the full catalog. Track the road to one hundred in ROADMAP.md.

Quickstart

git clone git@github.com:dimitris-di/lude-kit.git
cd lude-kit

# Claude Code skills (~/.claude/skills/)
./install/install-claude.sh

# Claude Code subagents (~/.claude/agents/)
./install/install-claude-agents.sh

# OpenAI Codex skills (~/.agents/skills/)
./install/install-codex.sh

# OpenAI Codex subagents as skills (~/.agents/skills/)
./install/install-codex-subagents.sh

# Slash commands (~/.claude/commands/ and ~/.codex/prompts/)
./install/install-claude-commands.sh
./install/install-codex-commands.sh

The scripts symlink, so future git pulls update your installed skills and subagents without reinstalling.

Custom install location:

CLAUDE_SKILLS_DIR=/path/to/claude/skills ./install/install-claude.sh
CLAUDE_AGENTS_DIR=/path/to/claude/agents ./install/install-claude-agents.sh
CODEX_SKILLS_DIR=/path/to/agents/skills  ./install/install-codex.sh
CODEX_SKILLS_DIR=/path/to/agents/skills  ./install/install-codex-subagents.sh

That's it. Start a conversation and the matchers pull the right skill when its triggers fire. Spawn named subagents with Agent(subagent_type: "architect", ...) and similar.

What's in the library

Seventy skills across three categories, organized so an orchestrator can scan and dispatch.

Category Count What it contains
personas/ 40 Senior grade role personas an orchestrator can dispatch to.
capabilities/ 10 Cross role capabilities focused on a single job.
stacks/ 20 Stack and technology specific experts.

Full catalog with one liners and links: skills/README.md.

The personas (40)

SDLC team (the people you'd staff a feature with): architect, tech lead, PM, UX, frontend, backend, DevOps/SRE, QA, security, tech writer.

Specialty roles: data engineer, ML engineer, data scientist, MLOps, mobile, embedded, game, blockchain, platform, developer advocate.

Industry verticals: fintech, healthcare, gov tech, edtech, ecommerce, media streaming, IoT fleet, automotive, compliance, logistics.

AI engineering: LLM app, AI agent, RAG, eval, fine tuning, voice AI, computer vision, recommender, model router, AI safety.

The capabilities (10)

Cross role power tools: code review, debugging, refactoring, performance engineering, incident command, API contract design, data modeling, migration planning, dependency auditing, postmortem authoring.

The stacks (20)

Frameworks and platforms: Rails, Django, Next.js, Kubernetes, Terraform.

Datastores: Postgres, Redis.

Clouds: AWS, GCP.

Mobile: Swift/iOS, Flutter, React Native.

Languages: Go, Rust, Python, TypeScript, Java, C#/.NET.

UI and testing: Tailwind, Playwright.

Subagents (30)

Skills load on demand inside a conversation when their triggers match. Subagents are named entry points you (or an orchestrator) explicitly dispatch via the Agent tool. They restrict tools, pin a model where it matters, and carry a system prompt that locks the right skill into the spawned conversation.

Lude Kit ships 30 curated subagents in subagents/, grouped three ways:

Specialists (10) mirror the highest leverage skills: architect, code-reviewer, security-reviewer, debugger, refactorer, perf-investigator, test-engineer, tech-writer, ic-coordinator, postmortem-writer.

Orchestrators (10) plan and dispatch multi skill flows: orchestrate-feature-build, orchestrate-incident-response, orchestrate-migration, orchestrate-launch, orchestrate-security-review, orchestrate-perf-investigation, orchestrate-refactor, orchestrate-ai-feature, orchestrate-new-service, orchestrate-bug-fix.

Library maintenance (10) keep the skill library healthy: skill-author-persona, skill-author-capability, skill-author-stack, skill-reviewer, skill-trigger-tightener, skill-deduplicator, skill-handoff-auditor, skill-freshness-checker, skill-catalog-updater, skill-eval-runner.

Install them with ./install/install-claude-agents.sh. After installation, dispatch with Agent(subagent_type: "architect", prompt: "design the billing service") and the subagent spawns with the right tools, model, and skill pre loaded.

Codex has no native subagent primitive equivalent to Claude Code's named dispatchable types. Each subagent is functionally a skill (description + system prompt), so ./install/install-codex-subagents.sh wraps every subagent as a Codex skill folder under ~/.agents/skills/<name>/SKILL.md. They auto activate via description matching on the Codex side instead of explicit dispatch.

Slash commands (10)

Pre written prompts you invoke with /<name> inside Claude Code or Codex. Each one dispatches the right specialist subagents in parallel and returns a synthesized verdict, so you do not have to retype the orchestration logic every time.

Command What it does
/review-macos-app Deep parallel review of a macOS app (Swift, sandboxing, perf, UX, App Store, OSS readiness).
/review-website Web app review (perf, a11y, security, SEO, framework idioms).
/review-api-service Backend API review (contract, schema, auth, observability, OSS readiness).
/review-mobile-app iOS / Android review (cross platform decisions, offline, push, store policies).
/review-cli-tool CLI review (UX, exit codes, signal handling, distribution channels).
/prep-oss-release OSS readiness sweep (LICENSE, README, SECURITY, CI, secrets check).
/security-audit Focused threat model + secure code review + supply chain.
/perf-audit Measure first, find dominant cost, propose budget and regression guard.
/a11y-audit WCAG 2.1 AA pass on any UI surface (web, native, mobile).
/lude-style Apply Lude Kit prose rules (em dash + hyphen sweep), identifiers preserved.

Install once with ./install/install-claude-commands.sh and ./install/install-codex-commands.sh. Then in a Claude Code or Codex session, type /review-macos-app (or any other) and the orchestration runs.

How a skill works

Each skill is a folder containing a single SKILL.md with YAML frontmatter and a Markdown body. The frontmatter holds the trigger description that matchers preload at startup. The body holds the brief that loads only when the description matches the user's intent. This is progressive disclosure, your context window stays cheap until the skill is actually needed.

skills/personas/staff-software-architect/
└── SKILL.md
---
name: staff-software-architect
description: >
  Use when designing a system, choosing a database / framework / cloud /
  message bus, writing an ADR or RFC, deciding build vs buy, planning
  capacity or scaling, reviewing an architecture diagram or proposal,
  sequencing a migration, or weighing technical trade-offs at the CTO
  level. Produces ADRs, RFCs, system diagrams, capacity plans. Not for
  implementation work, hands off to senior-backend-engineer /
  senior-frontend-engineer.
license: Apache-2.0
metadata:
  version: "1.0.0"
  category: persona
---

# Staff Software Architect

## Role
...

Every skill in this library follows the same nine section structure (Role → When to invoke → Operating principles → Workflow → Deliverables → Quality bar → Antipatterns → Handoffs → Quick reference) so an orchestrator can scan many of them quickly and reviewers know exactly what's missing.

The format follows the open Agent Skills specification, so the same folder runs on Claude Code and OpenAI Codex with no per platform duplication.

Multi agent composition

Lude Kit is designed for orchestrator style setups where a coordinating agent spawns subagents with specific roles. Every skill ends with a Handoffs section that names the partner skills for adjacent work, so the library composes:

User: "We want to add team SSO to the dashboard."

→ senior-product-manager       writes the one pager + success metric
→ staff-software-architect     picks SAML vs OIDC, writes the ADR
→ principal-security-engineer  threat models the auth flow
→ senior-ux-designer           flows sign in + invite + role switch
→ senior-frontend-engineer     implements the UI
→ senior-backend-engineer      implements endpoints + sessions
→ senior-qa-test-engineer      writes the regression suite
→ senior-devops-sre            rolls out behind a flag, sets SLOs
→ senior-technical-writer      writes admin doc + release notes

Each skill knows which skill to hand off to next. The library is a graph, not a list.

Roadmap

Batch Theme Status
1 SDLC personas (architect → writer) ✅ shipped
2 Capability skills (review, debug, refactor, perf, postmortem, …) ✅ shipped
3 Stack experts (Rails, Django, Next.js, K8s, Postgres, Terraform, Redis, AWS, GCP, iOS) ✅ shipped
4 Specialty personas (data, ML, MLOps, mobile, embedded, game, blockchain, platform, devrel) ✅ shipped
5 Industry verticals (fintech, healthcare, gov, edtech, ecommerce, streaming, IoT, automotive, compliance, logistics) ✅ shipped
6 AI engineering (LLM apps, agents, RAG, eval, fine tuning, voice, CV, recommender, gateway, safety) ✅ shipped
7 Language and framework stacks (Go, Rust, Python, TS, Java, .NET, Flutter, RN, Tailwind, Playwright) ✅ shipped
8 More capabilities (a11y, i18n, feature flags, chaos, FinOps, modernization, releases, on call, DX, docs) planned
9 AI infra (vector db, GPU infra, serving) + more verticals planned
10 Hardening pass: cross skill review, polish, 1.0 release prep planned

Target: 100 skills, then 1.0. Want to shape what comes next? Open a new-skill issue.

Contributing

PRs welcome. The bar is high and the process is opinionated. Read CONTRIBUTING.md before opening one.

Quick links:

Validation

A GitHub Actions workflow runs the linter on every PR. Run it locally with:

pip install pyyaml
python3 scripts/validate-skills.py

It checks every SKILL.md and every subagents/*.md for YAML validity, name/folder match, description budget (1024 chars), and no em dashes in the body. See SKILL_LINT.md.

Every skill in this repo is hand written, reviewed against the style guide, and verified against real prompts before merging. No autogenerated filler. If you submit a skill that does not change behavior in a useful way when activated, it will be declined with notes.

Who's behind this

Built and maintained by Dimitris Dimitriou (@dimitris-di). Questions, ideas, or pushback: open a Discussion or reach out at demetrisd25@gmail.com.

Acknowledgments

License

Apache-2.0. Use it, fork it, ship it.

About

A curated library of senior grade Agent Skills and subagents for Claude Code and OpenAI Codex. 70 skills, 30 dispatchable subagents, designed for multi agent orchestration.

Topics

Resources

License

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors