Vibe-coding is great for demos but terrible for production. That’s the core thesis of a new workflow guide for OpenCode published by developer Farishan on November 17, 2026. The argument is straightforward: the bottleneck isn’t the LLM’s capability, but the lack of structural guardrails around it. Without a defined spec and automated verification, agents produce code that looks right but fails in practice.

The Five-Step Loop

The proposed workflow enforces a rigid sequence: Spec, Implement, Verify, Review, and Ship. Each stage is gated, meaning you cannot progress to the next step until the current one passes. The 'Spec' phase requires three short filesβ€”SPEC.md, PLAN.md, and ACCEPTANCE.mdβ€”to define exactly what 'done' looks like before any code is written. This prevents the agent from hallucinating scope or losing track of requirements mid-session.

Verification as the Critical Gate

The most significant technical addition is the 'Verify' step, powered by a universal command called wf-verify. This tool detects the project type (whether it’s Go, Bun/Node, Rust, Python, or Make-based) and runs the appropriate build, lint, and test commands. It returns a binary pass/fail status. If wf-verify fails, the loop halts; you do not review or ship until the gate passes. This eliminates the 'works on my machine' excuse by standardizing the check across ecosystems.

Review and Ship Discipline

After verification, a distinct 'Review' pass checks the diff against the original SPEC.md to catch scope creep and security issues like hardcoded paths. The final 'Ship' step is mundane but mandatory: commit, tag, push, and update documentation. The author provides MIT-licensed templates and agent personas for these steps, along with a free core bundle and a $29 Pro version that includes additional diagnostic tools like sweep and doctor.

Key Takeaways

  • Structure beats raw model power: The workflow emphasizes that missing gates, not model limitations, cause vibe-coding failures.
  • Universal verification: The wf-verify command abstracts ecosystem-specific build/test commands into a single pass/fail gate.
  • Strict gating: You cannot proceed from Implement to Review if Verify fails, enforcing quality control before human attention is spent.
  • Open tooling: Core templates and prompts are MIT-licensed and available on GitHub, while advanced diagnostics are behind a paywall.

The Bottom Line

This is a necessary correction to the 'prompt and pray' approach many developers have adopted. If you’re using OpenCode for anything beyond toy projects, skipping the verify gate is professional negligence.