I build software for paying clients with Claude Code and Codex. This kit is that whole workflow, packed as files one script installs: the rules both engines follow, the review one engine gives the other's work, and the tests that must pass before anything merges. AI slop gets a second engine's review and has to pass your own tests before the merge script lets it through. Flip one switch and the script also refuses any commit without a saved review verdict.
For anyone who codes with Claude Code and Codex: developers, agencies, freelancers, founders. The rules send routine work to the cheaper models and keep the expensive reasoning for the places where a mistake costs more, so your Claude and Codex subscriptions go on the work that needs them. Built by a professional web developer who vibe codes for paying clients and would not merge a line without the review and the gate.
$29 for the first 30 sales, then $39. One payment in USD through Stripe. Download right after paying.
One rulebook both engines read, so Claude Code and Codex follow the same standards on every job: small files, clean commit messages, no secrets printed, no AI attribution in a client's repository, plain language in everything they write. Claude Code does the writing in its own git worktree. Codex reviews the result read-only and has to end with a verdict, saved under the commit it reviewed. Your own tests run on that same commit, and the merge script refuses to touch GitHub without a clean pass on it. Turn on the review switch and it also refuses a commit without the saved ACCEPT. It merges exactly the commit that was checked. Routine work goes to the cheaper models; the expensive reasoning is kept for design, hard debugging and review, where a miss costs the most.
It is not a course, a prompt pack or a hosted service. You get readable Bash and Python, the rule text both engines load, and GUIDE.md on why each piece exists and the mistake that produced it. No AI subscription, no server, no client code and no client text is included.
The scripts enforce what GUIDE.md describes, and the guide was written from daily client work, not from a demo. Every number in it comes from my own logs, with the date it was recorded. When the guide and a script disagree, read the script.
The installer previews each change and asks y/N before applying it. Buyer rule content is hidden; other previews hide lines that look like secrets.
Every file the installer replaces or creates goes into a private backup folder with a manifest. uninstall.sh walks that manifest in reverse and puts the old content back byte for byte.
Developers, agencies, freelancers, founders: if these two engines write code that other people depend on, the kit fits. Most of these should be true.
rules/
RULES.md holds the shared rules: which rule wins, writing, git and GitHub, secrets, code comments, tokens, long tasks and cleanup. claude.md and codex.md hold the few lines that differ. Claude Code imports the live files; sync.sh builds the Codex copy.
hooks/ + config/
ai-stop-hook runs the memory sweep when the module is installed, then plays a sound. ai-ding plays it on WSL, macOS and Linux with PulseAudio, and doubles as the Codex notify command. Settings snippets for both engines, an auto-mode environment template and efforts.md, which model and effort for which job.
skills/agent-harness/
Split work into pieces you can verify, isolate each worker, keep plans and progress in files, restart fresh instead of patching a tired context, and have results checked independently. Symlinked into both engines.
workflow/
cross-review.sh, gate.sh, merge-pr.sh, chain.sh, check-file-length.sh, check-commits.sh and local-ci.example.sh, plus worker rules, the review prompt and the plan, handoff, PR body and queue templates. WORKFLOW.md is the loop on one page.
memory/ (optional)
Capture rules, a Stop-hook sweep, an MCP client and a weekly review in which a second session verifies every change before it is written. Needs your own MCP memory server; none is included.
install.sh + uninstall.sh + GUIDE.md
The installer previews each change, asks per step, hides buyer rule content and secret-looking lines, and backs up everything it replaces. uninstall.sh restores from that backup. GUIDE.md explains each piece and the mistake that produced it.
Version 1.0.2. README.md inside the download lists every file and which engine each part needs.
Download the zip from your order page and unzip it. The kit includes README.md with its requirements and GUIDE.md with the reasons behind the workflow.
Run bash install.sh from the unzipped two-engines-kit folder. It previews the changes and asks y/N before applying each step.
Your existing rules stay active in both AI engines automatically. The installer warns if a CLI is missing and skips memory when no compatible server is configured. No moving, copying or editing files is required.
Restart Claude Code and Codex. If an engine asks you to approve the installed hooks, answer its prompts in /hooks. The included workflow examples are available for optional project customisation.
The download contains no hosted service, no memory server and no client code. README.md inside it repeats these requirements in full.
One-time payment in USD through Stripe. No subscription.
Everything listed above, GUIDE.md and version 1.0.2. Download right after paying. Buyers who give a GitHub username at checkout get read access to a private repository with updates. No release schedule is promised, and the version you have is yours to keep.
Sold by RayaStats. Read the licence and refund terms before you buy. Refunds cover a kit that does not work, reported within 7 days of purchase: if an installation fault cannot be fixed within 3 days of your report, you get a full refund. Not for a change of mind once the download has started.
Provided as is, without warranty. The full text is LICENSE.md in the download. A refund ends the licence and removes update access.
Requirements, licence, refunds and support, in plain words.
Anyone who uses Claude Code and Codex on work that other people depend on: developers, agencies, freelancers, founders. It helps to be comfortable reading a shell script before you run it. You bring your projects, your test commands and your own engine access.
Yes. The workflow is the two working together: Claude Code writes in its own worktree, Codex reviews read-only and leaves a saved verdict, and the merge script can be told to insist on that verdict. cross-review.sh calls the Codex CLI, the memory module's weekly review runs Claude Code, and merge-pr.sh and chain.sh need gh and git. README.md says which engine each part needs.
It installs the kit files, preserves your existing rules and imports in private files, points ~/.claude/CLAUDE.md at the shared rules and preserved content, builds ~/.codex/AGENTS.md from them, adds the Stop and Notification hooks and an effort setting to ~/.claude/settings.json, adds three keys to ~/.codex/config.toml only where missing, offers to remove saved Codex rules that allow push, merge, deploy, rm or drop, and adds .ai/ and .local-ci/ to your global git ignore. Each step previews its changes and asks; buyer rule content is hidden in previews. Every replaced or created file is backed up with a manifest, and uninstall.sh restores it byte for byte.
No. The memory module needs your own compatible HTTP MCP memory server. If none is configured, the installer skips it automatically; rerun the installer after registering one to add it. MCP is the protocol the engines use to talk to that server. Claude Code and Codex are separate products with their own subscriptions, and no subscription is included.
Yes, when the kit does not work. Report an installation fault within 7 days of purchase with your operating system and the error, secrets removed. If RayaStats cannot get it working within 3 days of your report, you get a full refund. A broken zip, missing files, a kit that does not match its description, or an accidental duplicate charge also qualify. If you have not ticked the immediate-supply box and have not downloaded, email support within 14 days of your payment for a full refund. The Download button on the order page works only after you tick a box asking for immediate supply, and once the download starts there is no refund for a change of mind. The terms give the full conditions and leave your statutory rights untouched.
Yes. Use and change it on any number of your own machines, for your work and your team's work, and keep the version you have for as long as you like. Do not resell, sublicense, redistribute or publish it, share it outside your team, or remove the licence. A refund ends the licence.
Buyers who give a GitHub username at checkout get read access to a private repository with updates. Access is removed after a refund. There is no promised release schedule, and the version you bought stays yours.
Email iliya@rayastats.com about installation errors, missing files or download trouble, with the email you used at checkout, your operating system and the error, secrets removed. Support covers getting the kit installed and running as described. It is a set of files, not a managed setup or consulting service.
One rulebook, a cross-engine review with a saved verdict, a local gate on the exact commit, and a merge script that refuses an ungated commit and, with the review switch on, an unreviewed one. Plain files, a reversible installer, and the guide on why each piece exists.
Questions first? Email iliya@rayastats.com