WongStack

Coding agents forget your decisions between sessions, and your process lives in chat and in people's heads. WongStack keeps the process, the plans, and the decisions in the repo, and gives agents a repeatable loop that writes down what each change teaches.

Each plan gets a review.html page like this. You read each change, see it drawn, and add notes before any code is written.

Start here

Make an empty folder and open it in your coding agent. Have a free Cloudflare account ready. Then paste this:

Read and follow
https://raw.githubusercontent.com/matthewwong525/WongStack/refs/heads/main/.agents/skills/wong-setup/SKILL.md
to install WongStack in this folder and walk me through the first workflow.

The agent asks for one Cloudflare token first, with the exact click path. Then it asks how you work, plans the install, applies it, and saves the result. You end with the workflow, a starter app online, and session memory. In a folder that already has files, setup stops and says so.

Claude Code and Codex get full support. Other agents, such as Cursor, can follow the workflow skills, because they are plain Markdown files. They get no hooks and no session memory.

What you get

• Process in the repo, not in chats, docs, and people's heads.
• One command for each stage of the work, from the first idea to the merge.
• A record written during the work. Plans, decisions, and lessons are written down as part of each step, not afterward.
• A reviewable package for each change, which your team inspects before it joins main.
• Memory across sessions. Each session starts with the facts that earlier sessions learned. The store is in your Cloudflare account, not in git.
• No lock-in to one agent. The durable part is the files in the repo.

The commands

/explore -> /plan -> /apply -> /save -> /continue -> /ship

You do not have to run each one. A command whose input is missing runs the one before it, so /apply plans first when there is no plan. The change loop owns the details.

Setup puts the app online on the Cloudflare stack: each change gets a preview link, and a merge deploys it.

How it compares

Checked against each project's README in September 2026.

Repository layout

Requirements

• A coding agent that edits files, runs shell commands, and asks questions.
• gh, signed in, and git. Setup creates the GitHub repo.
• Node.js 22, the version in .nvmrc.
• OpenSpec: npm install -g @fission-ai/[email protected].
• curl, for Cloudflare provisioning.
• A Cloudflare account (the free plan works) and one user token. Cloudflare is required, because session memory and hosting run there. The token stays in the git-ignored .env on your computer. SECURITY.md says what each token can do, and the credentials page has the click path.
• On Windows, run git config --global core.symlinks true and turn on Developer Mode before you clone. The skills folder links are symbolic links (why).

The setup prompt helps with missing pieces. Required tools says why each is needed.

Work from the source

Fork first, then clone your fork:

gh repo fork matthewwong525/WongStack --clone && cd WongStack

A clone of this repository sends /save pushes to a repository you cannot write to. In your fork, the commands work, because this repo uses WongStack itself. Session memory stays off: .agents/.wong-stack.json names the maintainer's store, and you have no token for it. The contributing guide says how to send a change back.

Learn more

• Wiki: the guide to WongStack's process, and the terms the agent may use.
• AI knowledge centers: the six principles behind WongStack.
• The change loop: how work moves from idea to shipped record.
• Working on WongStack: how to change the toolkit itself.
• Changelog: what changed between releases.

License

WongStack is released under the MIT License.