Show HN: Verification-first workflow plugin for Claude Code
github.comI built a set of open-source Claude Code plugins that add a structured define → execute → verify loop to agentic coding workflows.
*The problem*: Claude Code is capable, but on non-trivial tasks the default workflow is iterative — you prompt, review output, correct, re-prompt. Each cycle costs time and tokens, and the failure mode is accepting "looks right" code that breaks in edge cases you didn't think to check.
*What this does*: You run two commands:
- `/define` — Converts a task description into concrete acceptance criteria and invariants before any code is written. The agent explores the codebase, identifies constraints, and produces a manifest of what "done" means. - `/do` — Executes against the manifest, tracking progress per criterion. Verification runs automatically at the end — each acceptance criterion is checked as a separate pass/fail gate. If something fails, it fixes and re-verifies before declaring done.
That's it. Define what you want, then execute. The verification loop is built in.
*How it differs from Plan mode*: Claude Code's built-in Plan mode lets you review a plan before execution. But the plan is unstructured text — there's no formal criteria, no verification after execution, and no way to know if the output actually matches what you planned. This adds structured acceptance criteria that get verified automatically, so "done" means "all criteria passed," not "the agent stopped."
*How it differs from manual prompting*: Writing good prompts helps. But prompts describe intent — they don't verify outcomes. This separates the "what do I want" step from the "did I get it" step, with explicit checks for each.
Built with Claude Code's plugin system (agents, skills, hooks). No external dependencies.
No comments yet.