/ml-features

Make it your own.

In Claude Code, use the slash command and add your context. In Codex, select ml-features from the just-vibe skill picker, then send the same brief.

Version 0.11.0 also supports /jv ml-features, /just-vibe ml-features and /jv:ml-features in Claude. See shortcut setup and context examples.

What the agent does

1. Define row grain, entity/event identity, prediction instant, feature event time, source availability time, revision order and missing behavior. Specify timezone and exact window endpoints. Missing availability evidence blocks historical-validity claims.
2. For versioned sources, reconstruct the latest version actually available at each prediction instant before applying its event-time window. A later correction can move an event outside the window; filtering versions first can resurrect an obsolete value. Deduplicate by the documented event/version identity, scoped to the entity where required.
3. Fit learned transforms only on eligible training rows within each simulated fit or cross-validation fold. Preserve input ordering/identity and define empty, constant, missing and unseen-category behavior; use the same transformation semantics at serving time.
4. Verify exact time boundaries, mixed explicit offsets, late arrivals, corrections, duplicates, negative/zero values and no input mutation. Compare engineered features to an independent tiny example before proposing a usefulness experiment.
5. Only claim usefulness after a controlled baseline comparison on the chosen evaluation protocol and compute budget. A correct feature builder alone establishes neither predictive gain nor production readiness.

Inputs

• task, feature sources, availability timing, baseline, and evaluation protocol.

Optional context: scope, references, constraints, successCriteria, environment, mode, budget.

Scope

Reads

Predictable, reproducible feature engineering and controlled usefulness checks; implement/run when requested.

Writes

Inspect/plan: inspect or propose; save requested artifacts only. Apply: edit the requested local implementation and perform relevant bounded checks while preserving unrelated work. Live data changes, remote actions and paid jobs require their resolved target and existing session authorization.

Mode

Plan feature definitions when requested; apply for requested feature-pipeline code and bounded fixture checks. Actual experiments additionally require data access and resource limits.

Prerequisites

Task definition, dataset identity, field semantics, entity/time keys, and permission to inspect bounded data. Record prediction moment, label horizon, sampling, and provenance. Preserve held-out evaluation boundaries; no data upload, label alteration, or feature fitting across splits implicitly.

Expected output

• Feature/time/identity contract, implementation when requested, eligible fitting population, boundary and revision evidence, and separately measured usefulness results if any.

How the work is checked

• Historical features use only versions available at prediction time. Training transformations exclude held-out and ineligible rows, and preserve the documented empty/constant/missing behavior.

When to stop or clarify

• No test-set-driven feature selection. Unsupported availability timing blocks claims that a feature is deployable.

Handling missing context

Infer

Read prediction moment, label horizon, entity/time keys, split policy and dataset provenance from the task and manifests.

Assume

Use explicit synthetic examples for design when raw data is unavailable; do not infer missing labels or fit preprocessing across held-out boundaries.

Ask

Ask when unresolved label timing, grouping or target semantics would change the split/features; do not demand a full dataset to explain the method.

Technical guidance

Evidence

Define feature semantics, prediction-time availability, units, ordering and missing/unseen-value behavior.

Method

Build transformations in the training/inference pipeline and evaluate usefulness under the same split protocol.

Pitfall

Target encoding or imputation fitted outside the training fold contaminates validation; feature importance is not causal effect.

Check

Hand-compute a small example and exercise future-only input, unseen categories and missing values through both train and serving paths.

Situational decisions

When records can be corrected after their original event time: Use availability and revision semantics to reconstruct the visible version first; never join historical predictions to today’s final mutable table without qualification.

When no mature training rows remain after eligibility filtering: Use the documented empty-transform behavior or report the missing training prerequisite. Do not fit preprocessing on validation/test rows to avoid the empty case.

The coding agent follows this workflow using its available tools. Installation does not grant service access or guarantee an outcome. Read the compatibility notes.

Keep exploring