vaibhav-space/my-ai-coding-workflow

Production-grade AI coding instructions, curated skills, and a 3-stage review pipeline (Planning → Critical Review → Diff Audit) built after 4,500+ hours with AI agents. Agent-agnostic.

What it solves

AI coding agents often make costly mistakes: they over-engineer simple tasks, hallucinate APIs, quietly reformat files, and claim tests pass without actually running them. This project provides a set of global rules and review prompts that force agents to plan first, get the plan criticised, and then prove their work with real diff and test output before merging.

How it works

The repo contains one global rules file (CLAUDE.md) and three prompts: planning.md, critical-plan-review.md, and production-readiness-review.md. The workflow is:

  1. Type /plan (or paste the planning prompt) and describe a task. The agent inspects the codebase and produces an interactive plan page with tickable steps. Nothing is built yet.
  2. Once you approve the plan, open a new chat and run /review-plan (or paste the critical-plan-review prompt). A fresh context has to look at the plan without the bias of having written it, so it can red-team assumptions and find flaws.
  3. After the agent implements the approved plan, open another new chat and run /review-prod (or paste the production-readiness-review prompt). This reads the actual git diff and runs tests, demanding literal command output as evidence. It ends with a "Could not verify" list for anything it couldn't prove.

The rules in CLAUDE.md also bake in principles like YAGNI (no speculative abstractions), diff inspection over conversational summaries, deterministic arithmetic (no mental math for money), and explicit sign-off for risky changes. Optional helper files add hard locks: a pre-commit hook that blocks commits when tests fail, and deny rules for dangerous commands like forced pushes or .env reads.

Who it's for

Solo developers and small teams who let AI agents write real code and want proof—not promises—that changes are safe. The setup is tested for Claude Code on Linux, but the prompts are plain text and can be pasted into Gemini CLI or Cursor if you copy them from the prompts/ folder.

Highlights

  • A three-stage review workflow: plan, criticise the plan, audit the result—each in a fresh context to avoid self-confirmation bias.
  • Global CLAUDE.md rules that apply across every project without polluting individual repositories.
  • Curated open-source skill packs: Superpowers (process discipline), Ponytail (minimal code), and I Have ADHD (clear communication preferences).
  • Optional enforcement files: deny rules for risky shell commands and a pre-commit gate that blocks broken commits.

Related

  • Project
  • Project
  • Project
  • Project