AI Coding OS · by Lumina

Anyone can call an LLM.
Shipping one takes a harness.

The model is a commodity — everyone has it. What separates a demo from a production system is the harness: evaluation, gates, readiness scoring, observability.

This is that harness — plug-and-play skills (SKILL.md) for Claude Code, Cursor and Codex, grounded in the same evaluation and readiness discipline as SWE-bench, DORA and the OWASP LLM Top 10. Not prompts, not another skill repo.

Generating code was never the hard part.

On SWE-bench (the coding benchmark), swapping the model moves the score about 1 point — swapping the harness, about 22. The model is a commodity; the harness is where the work, and the moat, live: review, tests, gates, security.

  • Every skill has a usage boundary — no overlap, no guessing.
  • Quality gates decide promote / hold / rollback on evidence.
  • Anti-patterns and an output contract in every workflow.

Concretely: the harness, as skills you own.

what a skill is

A SKILL.md file

Markdown + frontmatter with a usage boundary, workflow, promote/hold/rollback gates, anti-patterns and an output contract. Structure, not vibes.

how it installs

Drop it in your skills folder

Copy into ~/.claude/skills (or your Cursor/Codex setup). No new tool, no CI rewiring, no server.

how you own it

A library, not a SaaS

Portable files you can keep, read and adapt. Access terms are confirmed directly while checkout is not yet public.

A curated library, not a prompt dump.

The coding, agent, operations and review skills used to build real products — plug-and-play in your agent of choice.

Agent engineering

Design → run → harden

Architecture, implementation, runtime reliability, red-teaming and standards so your agent rules work across tools.

Quality & gates

Ship on evidence

Readiness scoring, eval lifecycle, dogfooding loops and controlled prompt/policy experiments.

Operations & security

Production-grade

Observability/SRE, security guardrails, web security and retrieval reliability.

Review

Defend the codebase

Layered PR review, deep code-quality audits and end-to-end release orchestration.

The PR review skill is free.

Run it on a real pull request and inspect the methodology: layered review, quality gates and a clear merge verdict. Public download is not open yet; request the sample through the contact form.

free skill

pr-review

Structured, evidence-driven PR review against your project docs and quality gates. Works in Claude Code, Cursor and Codex.

Request the sample

Built on published research.

The evaluation, gates, release and security discipline in these skills is grounded in the field's published research and the standards of record — the same public work the courses cite, not invented for a landing page.

Stop prototyping. Start operating.

The library is prepared for Claude Code, Cursor and Codex. Checkout is not public yet; use the contact form to request access or the free sample.