Mcode Stack Agent Skill

<p align="center"> <img src="docs/assets/hero.png" alt="MCode Stack: the workflows and skills the MiniMax Code team uses to build MiniMax Code" width="100%"> </p> <p align="center"> <a href="LICENSE"><img alt="License: MIT" src="https://img.shields.io/badge/license-MIT-blue.svg"></a> <a href="https://github.com/MiniMax-AI/mcode-stack/actions/workflows/validate.yml"><img alt="validate" src="https://github.com/MiniMax-AI/mcode-stack/actions/workflows/validate.yml/badge.svg"></a> <a href="stacks/"><img alt="skills: stacks/" src="https://img.shields.io/badge/skills-stacks%2F-ff7a59.svg"></a> <a href="https://github.com/MiniMax-AI/minimax-code"><img alt="for MiniMax Code" src="https://img.shields.io/badge/for-MiniMax%20Code-5b6cff.svg"></a> </p> <p align="center"><b>English</b> · <a href="README_ZH.md">中文</a></p>

The workflows and skills the MiniMax Code team uses to build MiniMax Code. Open source, install and use.

This is not a prompt collection or a list of "best practices". It is what people on the team actually use every day: how to keep AI-written code at a high bar, how to decide whether a task should go to a subagent, how long to wait on a long command, how to keep long sessions from going dull, and what goes into context on every turn. Each piece is a SKILL.md skill or checklist that installs straight into MiniMax Code.

• Quick start
• What's inside
• How it works
• Articles
• Stacks
• Contributing

Quick start

Claude Code, one command. The command adds this marketplace and installs a whole stack. In a shell:

claude plugin install ronny --marketplace MiniMax-AI/mcode-stack

Inside a Claude Code session:

/plugin install ronny --marketplace MiniMax-AI/mcode-stack

Use jiachen, req-tpm or youfang in place of ronny for those stacks. From a clone of this repository, scripts/install.sh installs ronny (scripts/install.sh jiachen, scripts/install.sh req-tpm or scripts/install.sh youfang for the others) with claude or codex when either is on PATH, and otherwise copies that stack's skills into ~/.agents/skills/.

Older Claude Code can add the marketplace first, then install:

/plugin marketplace add MiniMax-AI/mcode-stack
/plugin install ronny@mcode-stack

Codex. Add the marketplace, then install from the Plugins Directory or the CLI:

codex plugin marketplace add MiniMax-AI/mcode-stack
codex plugin add ronny@mcode-stack

Substitute jiachen, req-tpm or youfang the same way. To take a single skill instead of the plugin, copy that skill's folder into ~/.agents/skills/. Codex and MiniMax Code both scan that directory; the copy steps are below.

One skill, for all your projects:

git clone https://github.com/MiniMax-AI/mcode-stack.git ~/mcode-stack
mkdir -p ~/.agents/skills
cp -R ~/mcode-stack/stacks/ronny/skills/long-command-check ~/.agents/skills/

Open a new session and type /skills in MiniMax Code to see it. Run it with /long-command-check, or let the agent load it from its description.

The whole stack, as a plugin:

mcode plugin marketplace list    # shows your local plugin directory

Copy the whole stacks/ronny directory into that directory, then open /plugins in MiniMax Code and press Ctrl+R to refresh. A local plugin is installed as soon as it is in that directory; press Space in the panel to enable or disable it.

Every stack, from the MiniMax Code marketplace:

This repository is also a plugin marketplace, so you can install a stack without copying files and pick up updates later. In MiniMax Code desktop, open Plugins, choose Add marketplace and enter MiniMax-AI/mcode-stack; each stack then shows up as one plugin you can install, update or uninstall. Adding marketplaces is rolling out gradually, so your version may not have it yet.

<details> <summary>More install options</summary>

MiniMax Code discovers skills in these directories (one subdirectory per skill, containing SKILL.md):

• For the current project only: .minimax/skills/, .claude/skills/ or .agents/skills/ in the project
• For all projects: ~/.agents/skills/ (~/.claude/skills/ and ~/.codex/skills/ are read too)

mcode plugin add installs from the official catalog only; it does not take a GitHub URL or a local path.

The skills use the common SKILL.md format, so most of them also work in other agents that support skills.

</details>

What's inside

<p align="center"> <img src="docs/assets/skills-map.png" alt="The skills in stacks/ronny, grouped by moment in the day, plus what stacks/jiachen adds" width="100%"> </p>

When	Skill	What it does
Plan & split	subagent-split-check	Decide whether work goes to a subagent or stays in the main loop
	engineering-small-prs	Split large features, refactors and migrations into small reviewable PRs
	engineering-change-descriptions	Write PR and commit descriptions that say what changed and why
Context & tools	tool-budget-audit	Measure what tools cost in context; trim, tool search or programmatic tool calling
	context-tax-audit	What gets injected into every request, and what it does to the prompt cache
Long sessions & commands	long-session-check	Drop, summarize or write to file; compact or start fresh; memory scoping
	handoff-doc	A hand-off doc so a new session can pick up the work
	long-command-check	Ten-point audit of long shell commands and background tasks
Review	independent-review	Review a diff from a clean context: only diff, requirement and test results
	engineering-code-review	Review a change for code health: design, functionality, complexity, tests, approval risk
	engineering-review-comments	Write clear, severity-labeled review comments that explain the reasoning
	engineering-review-feedback	Handle reviewer comments as the author: fix, clarify, resolve disagreements
Experiments & hotfixes	harness-experiment	Small eval, one change at a time, repeated runs, look at flipped tasks
	trajectory-review	Look at how the agent got its result, not only whether it passed
Owning agent work	agent-brief	Before delegating: stakes, hypothesis, scope, what "verified" means
	evidence-first-review	When an agent says done: evidence first, what checks miss, silent decisions
	review-capacity-check	Don't launch more agents than we can stand behind
	taste-to-checks	Turn a repeat correction into a test, lint, hook or CI step
	engineering-emergency-changes	Decide whether a hotfix is a real emergency and run expedited review safely

The table above is stacks/ronny. stacks/jiachen adds MCode Review, PR review that has to prove its findings: the same rules in your agent (pr-review) and in GitHub or GitLab CI, a verdict decided by a script, and a weekly loop (pr-review-tuning) that tunes its rules on your own PRs. It also adds lifting-loops, which finds the loops where you keep acting as a manual step for your agents and lifts you out of them, and making-content, which turns notes and material into Xiaohongshu image posts and short vertical videos.

stacks/req-tpm is an optional stack for requirement management. The person who raises a product or technical requirement acts as its TPM together with their own agent: challenge it at intake and fix a measurable expected outcome, get a priority decision from the owner, record the design decision, track it on a cadence and flag stale items, pass ship gates, and close only with evidence against the expected outcome. The body lives in a Feishu doc; status and flow live in Meegle (Feishu Project). It is v0 and needs the meegle and lark-cli CLIs plus a one-time look-up of your Meegle space's fields and states; see stacks/req-tpm/README.md.

stacks/youfang has magic-move-deck, which turns an architecture or flow walkthrough into a Keynote-style Magic Move slide deck: every slide is a diagram seen through a camera, and the boxes and titles two slides share morph between them. The agent shows one sample slide in each of four styles (hand-drawn, minimal, dark neon, neo-brutalist) and builds the deck in the one you pick, as a single offline HTML file.

A stack usually has three kinds of things:

Kind	What it is	How to use it
Skill	A directory with a SKILL.md	Invoke it in MiniMax Code with /<skill-name>, or let the agent load it from its description
Checklist	Also a skill: the agent walks your project item by item. A plain-text copy lives in checklists/	Same as above, or just read it
Workflow	A WORKFLOW.md describing how this person works with MiniMax Code every day	Read it

How it works

<p align="center"> <img src="docs/assets/how-it-works.png" alt="From an Engineering Notes article to a skill, installed into MiniMax Code and run on your project" width="100%"> </p>

AI writes code faster every month. What separates results now is the layer of engineering habits around the agent: when to make it stop and check its own work, who reviews, where state lives, what counts as a real experiment. Those habits are hard to follow from an article, so we turned them into skills and checklists. They pair with our MCode Engineering Notes column: the article explains the idea, this repository makes it runnable.

We also believe harness engineering is a craft that moves forward by experiment. Everything here will keep changing as we learn. Take it and change it. To suggest a change here, open an issue; see Contributing.

Articles

#	MCode Engineering Notes	Companion skills
01	More subagents isn't better: manage context first, then parallelize	independent-review, subagent-split-check, harness-experiment
02	Too many tools will eat your context: from tool search to PTC	tool-budget-audit
03	Keeping long sessions sharp: compaction vs. memory	long-session-check, handoff-doc
04	How long should one bash command wait: from foreground blocking to background tasks	long-command-check
05	One small field can break the whole prefix cache	context-tax-audit

Stacks

stacks/ronny/               Ronny's stack (MiniMax Code R&D lead), the first and most complete one
stacks/jiachen/             Jiachen's stack: MCode Review (PR review in your agent and in CI), getting out of the agent loop, notes into posts and videos
stacks/req-tpm/             Optional: the requester and their agent act as a requirement's TPM (Feishu doc body, Meegle status); v0
stacks/youfang/             Youfang's stack: architecture walkthroughs as Magic Move slide decks, in four styles
stacks/<name>/              One directory per team member; teammates are welcome to add theirs
notes/                      Which runnable piece goes with each Engineering Notes article
community/                  Stacks from outside the team, added by collaborators
templates/stack-template/   Template for a new stack

Every stack is also a MiniMax Code plugin package. See stacks/ronny, stacks/jiachen, stacks/req-tpm and stacks/youfang for what each one contains.

Contributing

Issues and discussion are welcome. For now, pull requests for code, documentation, and stacks are accepted only from repository collaborators. If you are not a collaborator, open an issue to discuss first.

We still want to hear how you work with MiniMax Code, or any other agent, day to day. Share that in an issue. Collaborators adding a stack:

1. Copy templates/stack-template/ to community/<your-id>/.
2. Write a WORKFLOW.md: how you use the agent every day, what you hand to it and what you keep an eye on yourself.
3. Add at least one skill or checklist you really use.
4. Run node scripts/check.mjs, then open a PR with the output and a short record of real usage.

Details in CONTRIBUTING.md. Using something from here in your own agent? Tell us with an I use this issue.

Relationship to MiniMax Code

MiniMax Code itself lives in MiniMax-AI/minimax-code. This repository only holds workflows and skills, no MiniMax Code source code.

License

MIT, except where a stack's NOTICE.md says otherwise. The engineering-* skills are adapted from Google Engineering Practices; the adapted material stays under CC-BY 3.0. See stacks/ronny/NOTICE.md.