ML Vibe-Coding Workflow Studio v2.4 Pinned

Staged prompt architectures & data-leakage guards for Cursor, Claude Code, and Copilot

● State synced
Project Definition
Development Archetype

How you instruct the AI determines context degradation and silent error rates.

1. Human Blueprint + Staged AI
Context reset per phase. Guarded splits, explicit test locks. Recommended.
2. All-in-One Megaprompt
Dump entire spec at start. Prone to silent fit_transform leakage.
3. Unconstrained Ad-hoc Chat
One runaway chat session. Severe context rot & metric gaming.
Data Leakage Guards
Leakage Prevention 100%
Context Drift Vulnerability Low (8%)
Recommended Thread Resets 3
Pipeline Nodes 6
Interactive ML Lifecycle DAG (Click node to inspect stage prompt)
Stage 2: Active

Stage 2: Data Hygiene & Leakage-Proof Splitting

Zero Leakage Enforced

Ensure train/validation/test sets are established prior to computing statistics, encodings, or missing value strategies.

AI BANNED ACTIONS (FAIL SAFES)
  • BAN Calling .fit_transform() on full dataframe
  • BAN Imputing missing targets with median
HUMAN VERIFICATION CHECKPOINT
  • CHECK Verify test indices are persisted as test_idx.parquet
  • CHECK Run shape assertion: train_X.shape[0] + test_X.shape[0] == raw.shape[0]
COPYABLE AI STAGE PROMPT
Loading stage prompt...