What is tokens-per-task optimization?

Tokens-per-task measures the total tokens consumed from the start to the completion of a task, including the 'invisible' cost of re-fetching information that was lost during poor context compression.

Why use anchored iterative summarization instead of simple truncation?

Simple truncation or opaque compression often loses critical details like file paths or error messages. Anchored iterative summarization uses structured sections (Intent, Decisions, Files Modified) to force the preservation of essential data.

When is the best time to trigger context compression?

The ideal trigger point is between 70-80% context utilization. This provides enough buffer to summarize effectively without waiting until the agent completely runs out of memory.

How do you evaluate if a compression strategy is working?

We use probe-based evaluation, which asks the agent specific questions about the compressed history regarding factual recall, artifact tracking, and decision reasoning to measure functional quality.

Context Compression Strategies

Name: Context Compression Strategies
Author: EricGrill

byEricGrill

•

生产力与工作流

Optimizes long-running AI agent sessions by implementing structured context compression to maintain technical accuracy and minimize token waste.

This skill provides a comprehensive framework for managing massive conversation histories and large-scale codebases within Claude's context window. It shifts the focus from simple token reduction to 'tokens-per-task' optimization, ensuring critical information like file paths, technical decisions, and artifact trails are preserved through structured summarization. By employing techniques like Anchored Iterative Summarization and probe-based evaluation, it helps developers build reliable agents that maintain context across million-token sessions without losing track of previous modifications or design decisions.

主要功能

01Three-phase workflow management for multi-million token systems

02Tokens-per-task optimization to reduce re-fetching costs

03Artifact trail integrity tracking for file modifications

042 GitHub stars

05Anchored Iterative Summarization for structured history retention

06Probe-based evaluation for functional compression quality

使用场景

01Summarizing conversation history in long-running debugging sessions

02Managing context for agents working on codebases exceeding 5 million tokens

03Building evaluation frameworks to test LLM retention after history truncation

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add ericgrill/agents-skills-plugins context-compression

For use in Claude.ai and ChatGPT

Download Skill