English · 中文
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.
The engineering skills now install as one plugin, mcode (display name
MCode-Agent): a single entry skill decides which of four stages you are in —
start, work, verify, wrap up — and dispatches the rest. The plugin also ships
four agents (mcode:strategist / mcode:deputy / mcode:worker /
mcode:verifier) with their models bound in the package, so they work the
moment it is installed.
The content stack mcode-content (display name Mcode-Content-Check) carries
making-content and the third-party skill lieflat-less-ai-tone.
Claude Code, one command. The command adds this marketplace and installs a whole stack. In a shell:
claude plugin install mcode --marketplace zhongbiran/mcode-stackInside a Claude Code session:
/plugin install mcode --marketplace zhongbiran/mcode-stackmcode is the engineering lifecycle workbench. Install mcode-content the same way
for the content pipeline. From a clone of this repository, scripts/install.sh installs
mcode (scripts/install.sh mcode-content for the other) 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 zhongbiran/mcode-stack
/plugin install mcode@mcode-stackCodex. Add the marketplace, then install from the Plugins Directory or the CLI:
codex plugin marketplace add zhongbiran/mcode-stack
codex plugin add mcode@mcode-stackSubstitute mcode-content 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/zhongbiran/mcode-stack.git ~/mcode-stack
mkdir -p ~/.agents/skills
cp -R ~/mcode-stack/stacks/mcode/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 directoryCopy the whole stacks/mcode 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 zhongbiran/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.
Moving to another computer (MiniMax Code desktop):
git clone https://github.com/zhongbiran/mcode-stack.gitCopy the two directories into the plugin directory (Windows:
%USERPROFILE%\.minimax\plugins\, macOS: ~/.minimax/plugins/):
mcode-stack/stacks/mcode -> .minimax/plugins/mcode
mcode-stack/stacks/mcode-content -> .minimax/plugins/mcode-content
Then open /plugins, press Ctrl+R to refresh, and type @mcode in a
conversation. For engineering work only, mcode alone is enough.
The plugin directory is not hot-reloaded: refresh or reopen a session after copying. The four agents (
mcode:strategist/mcode:deputy/mcode:worker/mcode:verifier) ship with the plugin, with their models bound in the package. Nothing to configure after install.
More install options
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.
| 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/mcode.
stacks/mcode-content holds
making-content, which
turns notes and material into Xiaohongshu image posts and short vertical videos,
and the third-party skill
lieflat-less-ai-tone,
which detects and minimally rewrites AI-sounding Chinese prose from two rounds of
measured corpora (CC BY-NC 4.0, non-commercial; see
stacks/mcode-content/NOTICE.md).
stacks/mcode also carries the requirement skills (req-intake, req-design,
req-prioritize, req-track, req-ship, req-outcome): whoever raises a
product or technical requirement acts as its TPM together with their own agent,
from challenging it at intake to closing it with evidence. The body lives in a
Feishu doc; status and flow live in Meegle (Feishu Project), so those skills
need the meegle and lark-cli CLIs plus a one-time look-up of your Meegle
space's fields and states.
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 |
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.
| # | 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/mcode/ The lifecycle workbench: one entry skill, four stages, the review family
stacks/mcode-content/ Content pipeline: notes into Xiaohongshu posts and short vertical videos
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/mcode and stacks/mcode-content
for what each one contains.
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:
- Copy
templates/stack-template/tocommunity/<your-id>/. - Write a
WORKFLOW.md: how you use the agent every day, what you hand to it and what you keep an eye on yourself. - Add at least one skill or checklist you really use.
- 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.
MiniMax Code itself lives in MiniMax-AI/minimax-code. This repository only holds workflows and skills, no MiniMax Code source code.
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/mcode/NOTICE.md. The third-party skill
lieflat-less-ai-tone in stacks/mcode-content is CC BY-NC 4.0
(attribution-noncommercial); commercial use needs a separate license. See
stacks/mcode-content/NOTICE.md.

