Jev Architect

An agent skill to help you decide what to build with Jev, where it could improve existing software, and what to test before committing to it.

Explore the examples · Install · Try a prompt

Jev answers questions with defined possible outcomes. For example, it could judge whether a message needs a reply. Your code could then add that message to a reply queue, with a person deciding what to say.

Jev Architect guides your coding agent through choosing and testing those uses. Bring a problem, an idea, or a codebase. You don't need existing software to begin.

• Explore a new idea. What useful tool could you build if checking every message or event were affordable? Start with who it helps and test that premise.
• Improve an existing workflow. Compare Jev with the approach you use today, including mistakes, waiting time, cost, and review work.

The answer can be “keep it in code.” A free-shipping rule based on cart total needs an exact comparison, not a model call.

What to expect

Ask for an explanation and get one at your level. Ask for ideas and the agent first clarifies the users, problem, and constraints. For a concrete design, it works out the context Jev needs, the question to ask, and what code should do when the answer is uncertain or the request fails.

The next step is a small experiment with a baseline and a reason to stop if Jev isn't useful. Suggested designs are hypotheses, not evidence of speed, savings, or accuracy.

Install

This repository uses the portable Agent Skills layout. Install it with your preferred skill manager or clone it into your agent’s skill directory.

For the skills.sh CLI, paste this command into your terminal and choose your agent:

npx skills add karanb192/jev-architect --skill jev-architect

Claude Code

With Claude Code installed, paste both commands into your terminal (Terminal, iTerm, or your editor's terminal). These are shell commands, not chat messages.

claude plugin marketplace add karanb192/jev-architect
claude plugin install jev-architect@jev-architect-marketplace

Run claude in your terminal to start a new session after installing. Inside its chat, paste a prompt from Try it. You can also invoke /jev-architect:jev-architect there with your request.

To update an existing installation, run these commands in your terminal, then start a fresh Claude Code session. Use the scope you originally installed in if it wasn't the default user scope.

claude plugin marketplace update jev-architect-marketplace
claude plugin update jev-architect@jev-architect-marketplace

See Claude Code's plugin documentation for installation scopes and reload behavior.

Codex can use the included plugin manifest or the portable skill folder.

Try it

Paste these prompts into your coding agent's chat, not your terminal.

Explore a new language-learning tool, even if you haven't built anything yet.

Use Jev Architect. I want to help adults practise spoken English between lessons. What could I build?

Help a coding agent notice when it is stuck.

Use Jev Architect. My coding agent keeps retrying fixes that don't work. Could Jev help it notice when it's stuck and ask me for help?

Explore a browser assistant for finding supplier invoices.

Use Jev Architect. I want a browser assistant that finds invoices across supplier websites with different layouts. Could Jev help it choose what to click?

Why does it sometimes browse?

Basic explanations use the skill's built-in concepts. Current prices, API details, and recommendations for a named tool require checking official documentation. If those sources aren't available, the recommendation should stay conditional. A source stating a capability is not proof that it works well on your data.

What it will not do

• Treat Jev as a replacement for every LLM call.
• Promise speed, cost, calibration, or quality without measuring the target workflow.
• Assume an integration is current without checking official documentation and the installed system.
• Treat confidence as permission to act, or valid output as proof of a correct decision.
• Put private paths, traces, hidden instructions, or credentials in a shareable report.

This repository provides instructions for your agent. Installing it does not connect to Jev or run a benchmark. Building an integration requires separate API access and evaluation. This is an independent project, not an official TypeSafe skill.

Repository layout

skills/jev-architect/    Portable Agent Skill source
.claude-plugin/          Claude Code plugin and marketplace metadata
.codex-plugin/           Codex plugin metadata
tests/behavior.md        Prompts and expected behavior for manual checks

Use the behavior checks to test the skill in your agent. They test the advice it gives, not Jev's model performance.

License

MIT