System Active

DeepSeek
I/O Delegation

Route cheap I/O tasks to DeepSeek V4 Pro. Keep reasoning on Claude. Extend your weekly limits from 3 days to a full week.

Days Before

→

Days After

The Problem

Most of what Claude does isn't thinking. It's I/O.

Reading files, generating boilerplate, summarizing logs — these tasks burn through your context window without requiring Claude's reasoning capabilities. You're paying premium tokens for grunt work.

Architecture

Two-tier delegation

Claude acts as the senior engineer: reviewing, deciding, debugging. DeepSeek handles the bulk I/O: reading files, generating scaffolding, summarizing.

🧠

Claude

Senior Engineer

Reasoning • Debugging • Architecture • Security

8k tokens (raw files)

400 tokens (summary)

⚡

DeepSeek V4 Pro

I/O Processor

File Reading • Boilerplate • Summarization

Tools

Three CLI commands

Simple bash scripts that call DeepSeek's API. No complex setup, no MCP servers.

📖

ask-deepseek

Bulk file reader. Send multiple files, get a concise summary. Perfect for understanding codebases without loading everything into context.

ask-deepseek "explain auth flow" src/auth/*.py

✍️

deepseek-write

Boilerplate generator. Generate tests, configs, documentation templates. Claude reviews and makes surgical edits.

deepseek-write "pytest tests for UserService"

🧹

extract-chat

Transcript cleaner. Strip tool calls and metadata from session logs. Output clean human/assistant dialogue for documentation.

extract-chat < session.log > docs/chat.md

Token Economics

95% reduction on I/O tasks

DeepSeek processes the raw content. Claude only sees the summary.

Without Delegation

8,000

tokens consumed

→

With Delegation

400

tokens consumed

↓ 95% Token Savings

Decision Matrix

When to delegate

Simple rules: I/O goes to DeepSeek, reasoning stays on Claude.

⚡ Use DeepSeek

✓ Reading 2+ files at once

✓ Any file over 200 lines

✓ Generating test scaffolding

✓ README / documentation templates

✓ Summarizing logs or command output

✓ Cleaning session transcripts

🧠 Keep on Claude

✗ Debugging (needs reasoning about state)

✗ Security analysis (needs judgment)

✗ Architecture decisions

✗ Surgical edits (< 50 lines)

✗ Single small file reads (< 200 lines)