No Dump — Development skill for Claude Code

🚫 no-dump

**Stop dumping full logs into Claude.**

A Claude skill that saves tokens during coding and Kaggle projects. Claude asks for **only the output or code segment it actually needs**, and sends back **only the code that changed**. No more full logs, full notebooks or full files.

Why

When you work on a project with Claude, it often says "paste the output" to check progress. Full outputs and full files:

• waste tokens
• bury the useful lines in noise
• fill the context window faster, so Claude forgets earlier parts of your project

This skill makes Claude ask for the smallest piece that answers its question.

Before / After

**Before**

Claude: Run the training and paste the output. *(you paste 400 lines of epoch logs)*

**After**

Claude: Run this and paste the 3 lines it prints:

import numpy as np
l, vl = history.history['loss'], history.history['val_loss']
print('epochs:', len(l), '| any NaN:', bool(np.isnan(l + vl).any()))
print('best val_loss:', round(min(vl), 4), 'at epoch', int(np.argmin(vl)) + 1)
print('final loss/val:', round(l[-1], 4), round(vl[-1], 4))

Features

• Minimal output requests: specific lines, summary check cells or yes/no questions instead of full logs
• Minimal code requests: asks for an outline first, then only the relevant function or cell
• Minimal code replies: "replace these lines with these" instead of resending whole files
• Recovery rule: if a snippet breaks something, Claude sends the full affected function once
• Works beyond Python: terminal patterns like tail and grep
• Kaggle-aware: submission checks, /kaggle/input paths, data checks

Installation

Claude app (web / desktop)

1. Download this repo as a ZIP, or zip the no-dump/ folder.
2. Go to Settings → Capabilities → Skills.
3. Upload the ZIP.

Claude Code

git clone https://github.com/dLounce/no-dump.git
cp -r no-dump/no-dump ~/.claude/skills/

Repository structure

no-d