Project
dibble
Open source
The engineering context I kept repeating to Claude and GPT, packaged once as portable skills and deterministic checks that travel through agent workflows, local CLI, and CI.
- Agent Skills
- CLI
- Codex
- Claude Code
- Node.js
The Repeated Work
I use Claude and GPT for different parts of software work, but I kept carrying the same instructions between them: stay on the design system, check a package before installing it, derive tool schemas from Zod, render the UI before calling it done, and keep summaries tied to evidence. Custom skills and instruction files helped, but copies drifted between tools. Rules that should return the same answer every time also needed a deterministic layer. dibble puts that guidance and those checks in one versioned catalog.
One Working Loop
The 11 plugins follow one loop: trust the inputs, guide the work, and verify the result. install-gate checks an unfamiliar dependency before it enters the project. tokenlock keeps the implementation on the design system. design-verify asks for rendered evidence before the work is accepted. Each plugin stands alone, but the catalog covers the full path from trusted input to verified output.
The Contract Travels
Every plugin ships its working context as a plain Agent Skill. Claude Code and Codex have native marketplace metadata, while the deterministic layer is a set of zero-dependency Node scripts available through npx dibble and individual CLI bins. tokenlock and install-gate also package command hooks for both hosts. Codex requires the user to inspect and trust each hook definition before it runs, and Claude-only slash commands remain documented as host-specific. The model can change. The engineering contract stays put.
Judgment and Enforcement
The boundary is deliberate. Skills carry judgment and context. Scripts handle rules that should return the same answer regardless of which model or person produced the work. The same checker can sit behind a hook, run from a terminal, and fail a build in CI. Several of the design-system plugins came from leading engineering for a system used by 1000+ engineers and designers. plugin-inspector applies the same rule back to the catalog by validating both marketplace formats and reporting the authority each plugin declares.
Current Quality Bar
Release 1.2.1 is published to npm. CI runs 148 tests, enforces a 95% line-coverage floor with focused branch floors for the security-oriented checkers, and evaluates every skill against activation, missing-context, and unsafe-edge cases. The catalog also lints its own prose, audits every README, validates both plugin formats, and performs a package dry run. External adoption is still the evidence it needs next.
Stack
Node.js · Agent Skills · Claude Code and Codex plugin manifests · zero-dependency CLI checkers · GitHub Actions · semantic-release