- Home
- Skills
- Data & Databases
- ML Spec Author: Safe JSON ML Specs from Plain English
ML Spec Author: Safe JSON ML Specs from Plain English
Turns a plain-English ML request into a validated, code-free Open MLSpec JSON spec, with data-health checks and a plain-English explanation.
Free
Works with the AI Agents you already use
ML Spec Author: Safe JSON ML Specs from Plain English
Example session with this skill installed
I've got our customer export, one row per customer with a churned column. Set up a model that predicts who will churn. Don't train anything yet, I just want a spec our ML platform can run.
- Read your context and instructions
- Compiled the ml spec author
- Generated the data export
Plan: write churn.mlspec.json, a spec that predicts churned (yes/no) from plan, monthly spend, support tickets and recent logins. This is a classification task with a stratified 80/20 split, comparing logistic regression with a random forest on AUC.
1. What the data looks like (mlspec.py inspect customers.csv --target churned)
| Column | Type | Role |
|---|---|---|
| customer_id | identifier | identifier |
| signup_date | datetime | ignore |
| plan | categorical (basic, pro, team) | feature |
| monthly_spend | numeric, min 0 | feature |
| support_tickets | numeric, min 0 | feature |
| logins_last_30d | numeric, min 0 | feature |
| churned | categorical (no, yes) | target |
500 rows, no nulls, no health warnings. signup_date is set to ignore: a raw date would let the model learn the calendar rather than behaviour. Add a tenure feature later if you want it.
2. The spec (churn.mlspec.json, written next to your CSV)
- Dataset:
customers.csv, withcustomer_idunique andchurnednever null. - Split: stratified on
churned, 80% train and 20% test, seed 42.
Models: sklearn.linear_model.LogisticRegression 1.5.2 (class_weight: balanced) and sklearn.ensemble.RandomForestClassifier 1.5.2 (300 trees, balanced).
- Metrics:
auc,pr_auc(useful if churners are rare) andf1. - Workflow:
fit_logregandfit_forest, thenpick_bestby AUC, thenwhyto explain the forest.
3. Validation
The first draft gave signup_date the role metadata, and the validator refused it:
"valid": false
datasets.customers.schema.columns.1.role: 'metadata' is not one of ['feature', 'target', 'identifier', 'timestamp', 'group', 'weight', 'ignore', ...]
After changing the role to ignore:
{ "valid": true, "errors": [], "advisories": [] }
Next steps
- Nothing has been trained, so there are no accuracy numbers yet. Run the spec on an MLSpec-compatible runtime to get them.
- If churners are under about 10% of rows, compare models on
pr_aucinstead ofauc. - Want a forecast of churn by month instead? Ask, and I'll write a
forecastingspec with a time-based split.
customers.csv
CSV · data export
Example file from a real run - the skill writes it into your workspace.
Connects securely to your tools. The creator never sees your data.
About this skill
Describe what you want to predict or forecast, and your agent writes a declarative Open MLSpec JSON spec instead of training code: dataset schema, split, features, model, metrics and workflow steps.
What it does
- Profiles your CSV first (
mlspec.py inspect): proposes column types and roles, the target and the task, and warns about leakage, class imbalance or too few rows. Samples from columns that look like personal data are withheld. - Writes the spec using only registered model families, metrics and operations at pinned versions. Nothing is invented.
- Validates it (
mlspec.py validate) against the bundled JSON schema, a safety scan, the plugin registry and the workflow rules. Refusals go back to the agent, which fixes them. A spec is never called valid unless the script said so. - Explains it in plain English: what will be predicted, how the data is split, which models are compared and why.
Covers classification, regression, time-series forecasting and exploratory charts. Four example specs are included.
Why it is safe
- Declarative only: no Python, shell or SQL is written or run to train anything, and no code goes into a spec.
- No accuracy numbers are ever claimed, because nothing is trained.
- Your data stays local, and database credentials are referenced by environment-variable name, never read or printed.
- Python 3.10+ standard library only. No network access.
To train, run the validated spec on an MLSpec-compatible runtime.
Writes and validates specs only; it does not train, score or forecast, and it never reports accuracy. Covers classification, regression, forecasting and exploratory charts, not clustering or custom losses. CSV inspection loads the whole file into memory. Requires Python 3.10+.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 9 days ago
- Free with an Agensi account
Requires
Permissions
Runs the bundled scripts/mlspec.py locally to profile a CSV the user names and to validate the spec. Reads only files the user points to, writes the spec next to the data after asking, and makes no network requests.