jev-architect (BenjaminPolge)
Skill that makes Claude Code and Codex split AI features into Jev, LLM or plain-code steps before writing code.
Overview
jev-architect by Benjamin Polge is a skill for Claude Code and Codex that asks one question at the architecture step: does this step really need a generative LLM? When asked to build an AI feature, coding agents tend to use an LLM call for everything, including parts that are closed decisions such as classifying a ticket, scoring a lead, routing a request or answering yes or no. The skill splits a feature into atomic steps and assigns each to Jev, an LLM or plain code, comparing on eight axes: functional fit, cost, latency, volume, determinism, need for free generation, need for reasoning and integration cost.
It proposes an architecture, for example team routing on Jev Choice, urgency on Jev Score, reply drafting on an LLM and deduplication in code, and asks the user to choose. It also flags LLM calls in existing codebases whose output is already a closed decision. It does not call Jev, add runtime dependencies, rewrite code on its own or push Jev everywhere; once approved, the agent reads the official docs. It is Apache-2.0 licensed.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Other | Should this step run on Jev, an LLM or plain code? | Jev · LLM · plain code |
How it works
- User asks the agent to build an AI feature
- Skill splits the feature into atomic steps
- Each step is assessed on eight axes
- Agent proposes a Jev/LLM/code split for approval
- Agent consults official docs to implement
Strengths
- Prevents LLMs being used as classifiers
- No runtime dependency added
- Works on new and existing codebases
Caveats
- Advisory only; quality depends on the agent's analysis
At a glance
- Type
- Project
- Status
- beta
- Runtime
- agent skill
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Makes Claude Code and Codex ask whether a step needs a generative LLM at all — or whether it belongs on Jev, TypeSafe's System One model. Architecture arbitrage before the code is written.
hellogumbo/awesome-jev
Frequently asked questions
Does the skill call Jev?
No. It only proposes architectures; implementation is up to you.
What axes does it compare?
Functional fit, cost, latency, volume, determinism, need for generation, need for reasoning and integration cost.
Does it work on existing code?
Yes, it flags LLM calls whose output is already a closed decision.
Related projects
Portuguese agent skill teaching Claude Code, Cursor and Codex to use Jev correctly, bundled with the full raw research behind each claim.
Agent skillproductionAgent skill that has Jev pick the few relevant lines from big logs before your coding agent reads them, with a reversible reduce and ledger.
ProjectbetaPythonReadable reference skill for Claude Code and Codex where Jev picks from a numbered table of page controls, with an explainer site.
Agent skillprototypeJavaScriptLocal MCP bridge letting Codex hand drafting to Gemini via a Google AI Pro quota, with Jev checking drafts against explicit criteria.
MCP serverbetaChinese-language agent skill that offloads classify, screen, score and compliance checks to Jev via OpenCode Zen's free tier.
PluginbetaPythonScores installed Claude Code skills against the current project with Jev and hides irrelevant ones via skillOverrides.
Agent skillbetaJavaScript