GitHub repo packages a 3-stage AI code review workflow
A developer published the CLAUDE.md rules and three prompts he uses to stop AI agents from writing code nobody checked.
A rules file and three prompts, published as a repo
On 2026-11-14 a GitHub user named vaibhav-space created my-ai-coding-workflow, a repository that ships an AI coding review workflow rather than an app. By 2026-11-20 it had 304 stars and 45 forks, all collected in the six days since it was created. The main language is HTML and the topics are agentic-workflows, ai-agents, claude-code, code-review and developer-tools. The README frames the project as production-grade instructions, curated skills and a three-stage pipeline that runs planning, critical review and a diff audit. The repository announcement is the only source for the setup described here.
What is actually in the repo
The author is direct about scope. This is not a framework and not an app. It is one global rules file, CLAUDE.md, plus three prompts and some optional helpers. The rules are instructions an agent follows, not enforced locks. The README also states there is no public example project built with the setup yet.
The prompts live in prompts/: planning, critical-plan-review and production-readiness-review. The repo bundles curated skills with upstream attribution to Superpowers by Jesse Vincent, Ponytail by Dietrich Gebert and I Have ADHD by ayghri.
Setup for Claude Code needs bash and is tested on Linux. A script, install-commands.sh, adds three slash commands (/plan, /review-plan, /review-prod) and copies a plan page template to ~/.claude/plan-template.html. It skips files that already exist, takes --force to replace them, and keeps a .bak copy. People on Gemini CLI or Cursor are told to copy the prompts by hand instead.
How the three stages work
Stage one is planning. The agent inspects the repository, maps requirements to acceptance criteria and defines how each criterion will be verified. A hard instruction tells it to stop once the plan is published and wait for approval before touching implementation code.
Stage two is critical plan review, run in a new chat or subagent so the reviewer did not write the plan. It is asked to behave like an independent senior engineer looking for race conditions, missing error boundaries, timeout policies, idempotency risks and over-engineering. The author suggests running it on a different model from the one that wrote the plan, since different models miss different things.
Stage three is production readiness review, again in a fresh context. It audits raw git state: git status, git diff, git diff --cached and untracked files. Each check needs literal command output pasted in, or the word BLOCKED. New tests must be shown failing first.
The README also routes work by size. Small spikes get one cheap probe command and a short finding. Bounded tasks of one to three files run directly. Architectural or breaking changes go through the full pipeline.
What it changes for people building apps
The prompts are the least interesting part. The routing and the evidence rule are the substance. Instead of asking an agent to summarise its own work, the workflow asks for pasted command output, which is harder to fake than prose. The author's ten principles include keeping models away from mental arithmetic on currency and tax, and requiring human sign-off for schema migrations, public API contract changes, new third-party dependencies and destructive shell commands.
Limits and open questions
The 4,500+ hours figure is the author's own estimate, roughly six productive hours a day over a little more than two years, and the README says it is not tracked telemetry. The author does not claim equal time on the tools he names: Cursor, Claude Code CLI, Antigravity IDE and Gemini CLI, Codex, OpenCode and Amazon Q. The gates inside CLAUDE.md are behavioural instructions rather than mechanically enforced boundaries, and the README points to the coding tool's own permission settings for real enforcement. Without a public example project, the results rest on the author's account alone. The source text supplied also cuts off partway through the third review stage, so the full list of its checks is not covered here.
What this means for building apps with AI
The split between planning, critique and verification is portable even if you never clone the repo. Anyone describing an app in plain language and letting a model write the code faces the same problem: the agent grades its own homework. Draftship's guide to writing effective prompts covers how to phrase a request so the result is checkable, which is the same instinct this repository builds its pipeline around. Draftship is not affiliated with the project and does not bundle it.
Primary sources
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.