Personal project · 04
just-vibe
A toolkit for coding agents that turns everyday requests into focused engineering workflows, with reusable skills, project memory, and evidence-based verification.
Step through a mobile-menu fix, then change the check result to see the task return for revision. An illustrative workflow, not a live agent run. All portfolio information is available in the page text.
Request · Fix the mobile menu. Keep the existing design.
The work
From a request to work you can verify
A little structure for open-ended work
Coding agents can do a great deal, but a useful result still depends on choosing the right approach, keeping the user's constraints in view, and checking what actually changed. I built just-vibe to make those engineering habits reusable across everyday requests.
The toolkit installs into Codex and Claude Code. A user can describe an outcome, such as fixing a mobile menu, or choose a focused workflow explicitly. The coding agent carries out the work with the tools available in its current environment.
Designing the workflows
Each workflow defines when it fits, what evidence to inspect, how to proceed, and how to verify the result. Recovery guidance handles missing prerequisites and failed checks. Engineering profiles add priorities for different kinds of work, from frontend development to machine learning and application security.
The published v0.12.0 package includes 218 canonical workflows and 112 engineering profiles. Aliases share one implementation. A canonical catalog generates the host-facing skills and reference material, keeping instructions consistent across the two hosts.
Context that survives the next request
Trusted native hooks can select relevant workflows from an ordinary request and restore selected task context. Project instructions, preferences, and checkpoints help retain decisions through longer sessions. Explicit corrections can refine routing and workflow guidance; a passing test or silence is not treated as a user preference.
Memory tools preserve instruction provenance and check for concurrent file changes before writing. The goal is to keep the next step grounded in what the user has actually asked for, including constraints and corrections from earlier in the conversation.
Making results reviewable
Runtime helpers connect requirements to checks, screenshots, and attributed review. Evidence is tied to the files it describes, so later changes can make an earlier result stale. Alternative implementations can be compared in separate worktrees and the selected result applied with an undo record.
The exhibit shows a small illustrative example: a mobile-menu request selects a responsive-UI workflow, moves into implementation, and reaches verification. Change the check result to see the task return for revision. It visualizes the process; it does not execute a coding agent in your browser.
Packaging and testing
I built a dependency-free Node.js installer with setup, diagnostics, update, and removal commands. It checks the existing installation and host capabilities before making changes. The documentation website uses Astro and draws its command catalog from the same source as the plugin.
Validation covers generated-file consistency, command contracts, installation behavior, context preservation, and controlled implementation exercises. Native-host tests use isolated configuration directories. Browser checks exercise the website and supported interactive workflows. Those checks establish their tested scope, rather than promising that every workflow has been validated against every service or model.
Current scope
The npm package is MIT licensed. Installing it supplies workflows and utilities; it does not grant service permissions or provide missing credentials. Host tools and trust settings determine what a workflow can run. The source project continues to develop additional runtime capabilities separately from the published release.