Guide · Agent skill
How to use the ML Engineering Workflow agent skill (mle-workflow)
Updated · written with AI from the project's own documentation and checked against our catalogue data
The mle-workflow skill is a packaged set of instructions (a SKILL.md file) that tells an AI coding agent how to handle machine learning work as a production system rather than a one-off notebook. It covers data contracts, reproducible training, model evaluation gates, deployment artifacts, monitoring and rollback. This guide is for engineers who have never installed an agent skill; by the end you will know what the skill changes, how to install it, how to ask for it in plain language, and what to check before relying on it.
ML Engineering Workflowaffaan-m/ECC · MIT Download the skill Quick start
- 1Download the ZIP. Download the mle-workflow skill ZIP from this page.
- 2Unzip the skill. Unzip it so that SKILL.md is inside the mle-workflow folder.
- 3Place it in a skills folder. Move the folder into ~/.claude/skills/ for all projects, or into the project's .claude/skills/ for one repository.
- 4Or upload in the Claude apps. In the Claude apps, open Settings → Capabilities → Skills and upload the skill.
- 5Describe your ML task. Ask for the work in plain language and name the production concern, such as leakage or rollback.
- 6Provide project context. Share labels, traffic pattern, monitoring owner and existing pipeline paths so assumptions are not guessed.
- 7Review the output. Check the agent's contracts, tests and proposed commands before merging or running anything.
What is an agent skill, and what does mle-workflow make the agent do?
An agent skill is a folder with a SKILL.md file that gives a compatible AI agent extra instructions it loads when a task matches the skill's description. This one is named mle-workflow and is described as a production machine-learning engineering workflow for data contracts, reproducible training, model evaluation, deployment, monitoring, and rollback.
When it is active, the agent is directed to make model work explicit: define data contracts, make training repeatable, set measurable quality gates, produce deployable artifacts, and plan operational monitoring. The skill says it is meant to be scaled to the project and should not force heavy MLOps tooling where a data contract, baseline, eval script and rollback note would be enough.
- Checks data contracts, label and feature timing, and leakage risks
- Looks for reproducibility in training and consistent training and serving logic
- Plans evaluation gates, promotion criteria and rollback paths
- Points to related skills for testing, deployment and security
Install the mle-workflow skill
Download the ZIP for the skill from this page and unzip it. The skill folder must contain SKILL.md. Then put it in one of two places: your user-level skills folder at ~/.claude/skills/, or a project's .claude/skills/ folder if you want it to apply only to that repository.
If you use the Claude apps instead of a terminal, you can upload the skill under Settings → Capabilities → Skills.
- User-wide: unzip into ~/.claude/skills/
- Single project: unzip into the project's .claude/skills/ folder
- Claude apps: Settings → Capabilities → Skills, then upload
Trigger the skill with natural requests
You do not need a special command. The agent decides to use the skill when your request matches its description, so describe the ML task in plain words. Mention the production concern you care about, such as leakage, reproducibility, rollout or monitoring.
The skill's own activation list covers planning or reviewing a production ML feature, converting notebook code into a reusable pipeline, designing promotion criteria, debugging drift or training/serving mismatch, and adding monitoring or canary rollout.
- "Review this notebook and turn it into a reusable training pipeline with an evaluation step."
- "Check this feature code for label leakage and train/serve mismatch."
- "Plan promotion criteria and a rollback path for the new ranking model."
- "Add shadow traffic and post-deploy monitoring for this classifier."
What to provide so the agent can do useful work
The skill asks the agent to make assumptions explicit when information is missing. Give it as much project context as you can: whether labels exist, whether outcomes are delayed, how slices are defined, what traffic the model gets, and who owns monitoring. If you skip these, the agent should state its assumptions rather than invent them.
The skill also expects the agent to read existing code first. It recommends finding current training, feature, serving, evaluation and monitoring paths before introducing a parallel ML stack.
- Point the agent at the existing repository and notebooks
- Say whether the model is batch or online, and what it predicts
- Describe labels, delayed outcomes and any slice definitions
- State what monitoring or rollout tooling already exists
Requirements and tools
The skill is written for Claude Code or a SKILL.md-compatible agent. Its declared tools are Read, Write, Edit, Bash, Grep and Glob. The source also lists optional related ECC skills, such as python-testing, eval-harness and deployment-patterns, which deepen specific areas.
Most guidance assumes project infrastructure already exists, so it is a poor fit for someone without an ML codebase.
- Agent: Claude Code or a SKILL.md-compatible agent
- Tools: Read, Write, Edit, Bash, Grep, Glob
- Optional companion skills: python-testing, eval-harness, deployment-patterns
Safety notes: scripts and network
The skill itself does not run scripts and does not use the network, according to its listing. It is a set of written instructions. However, the agent it guides has Bash access, so review any commands it proposes before running them, especially those touching datasets, model artifacts or production systems.
The skill also asks for checks on secrets, PII, unsafe deserialization and supply-chain risk in artifacts, notebooks, prompts, datasets and logs. Treat those checks as part of your review, not a substitute for it.
- Listed as: runs scripts no, uses network no
- Review agent-proposed Bash commands before approving them
- Watch for secrets and PII in notebooks, datasets and logs
- Be cautious with model artifacts that use unsafe deserialization
Tips for getting good results
Ask for scope before implementation. The skill tells the agent to use only the lanes that fit your system, so say which parts matter now, such as a data contract or a rollback note, rather than a full platform.
Use the skill for review as well as building. Its guidance on leakage, reproducibility and promotion risk is useful when you paste in a pull request or an existing pipeline.
- Start with a small deliverable, such as a data contract and a baseline
- Ask for tests on feature transforms, metric calculations and inference schemas
- Ask the agent to record each production bug as a regression test
What people use it for
- Converting a notebook model into a reusable training and evaluation pipeline
- Reviewing an ML pull request for label leakage, drift risk or reproducibility gaps
- Planning a canary rollout or shadow traffic test for a new model version
- Writing promotion criteria and a rollback note for a ranking or classifier model
Frequently asked questions
How do I install the mle-workflow skill?
Download the ZIP from this page and unzip it so SKILL.md sits in a mle-workflow folder. Put that folder in ~/.claude/skills/ for all projects, or in a project's .claude/skills/ folder for one repository. In the Claude apps, you can instead upload it under Settings → Capabilities → Skills.
What do I need to run mle-workflow?
You need Claude Code or another SKILL.md-compatible agent with Read, Write, Edit, Bash, Grep and Glob tools. Optional companion skills such as python-testing, eval-harness and deployment-patterns add depth for specific work. It is most useful when you already have an ML codebase.
Is mle-workflow free?
The skill is listed under the MIT licence. This page does not state any price, so check the licence file in the source repository for the terms that apply.
Does mle-workflow train models or run experiments for me?
No. It is a guidance workflow that shapes how the agent plans, reviews and hardens ML work. It is not an ML library and does not train models or run experiments by itself.
Does the skill run scripts or access the internet?
According to its listing, it does not run scripts and does not use the network. The agent using it has Bash access, so you should still review any commands it proposes.
Why isn't the skill just a full MLOps setup?
The skill is designed to scale to the project. It warns against adding heavy machinery when a data contract, baseline, eval script and rollback note would be enough, so it is suited to teams that want rigour without a large platform.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use mle-workflow if you are turning notebook models into pipelines, reviewing ML changes for leakage or reproducibility, or planning rollouts and monitoring with an existing codebase. Skip it for simple one-off data exploration, for expecting the agent to train models by itself, or if you have no ML project to point it at.
Download the skill