What is tokens-per-task optimization?

Tokens-per-task is a metric that measures the total tokens consumed from the start to the completion of a task, including the hidden costs of re-fetching information if compression was too aggressive.

What is probe-based evaluation in context engineering?

Probe-based evaluation measures functional quality by asking the agent specific questions after compression—such as 'What was the original error message?'—to verify that critical technical details were actually preserved.

When is the best time to trigger context compression?

Ideally, compression should be triggered when context utilization reaches 70-80%. This ensures there is enough 'breathing room' to generate a high-quality summary before the window is completely exhausted.

How does Anchored Iterative Summarization differ from standard summarization?

Unlike standard summarization which regenerates the entire summary, Anchored Iterative Summarization uses explicit sections (like Decisions Made and Files Modified) and only merges new information into the existing structure to prevent silent information loss.

Context Compression Expert

Name: Context Compression Expert
Author: Kalyanikhandare29

byKalyanikhandare29

0•

Productivity & Workflow

Optimizes AI agent performance by implementing structured summarization and token-efficient context management for long-running sessions.

The Context Compression Expert skill provides advanced methodologies for managing massive conversation histories and large-scale codebases within AI context windows. It shifts the focus from simple token reduction to 'tokens-per-task' optimization, ensuring critical information like file modifications, architectural decisions, and artifact trails are preserved during truncation. By utilizing techniques such as Anchored Iterative Summarization and probe-based functional evaluation, it helps developers maintain agent continuity in complex, multi-step engineering tasks without losing technical accuracy or suffering from information drift.

Key Features

01Phase-specific workflows for managing codebases exceeding 5M+ tokens

02Structured artifact trail tracking to prevent loss of file modification history

030 GitHub stars

04Probe-based evaluation frameworks to measure functional compression quality

05Anchored Iterative Summarization to maintain structured, persistent session states

06Tokens-per-task optimization strategies to minimize expensive information re-fetching

Use Cases

01Debugging large systems where the agent must remember historical file changes and decisions

02Managing long-running agent sessions that exceed standard context window limits

03Summarizing complex codebase exploration into actionable implementation plans

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add kalyanikhandare29/agent-skills-for-context-engineering context-compression

For use in Claude.ai and ChatGPT

Download Skill

Key Features

01Phase-specific workflows for managing codebases exceeding 5M+ tokens

02Structured artifact trail tracking to prevent loss of file modification history

030 GitHub stars

04Probe-based evaluation frameworks to measure functional compression quality

05Anchored Iterative Summarization to maintain structured, persistent session states

06Tokens-per-task optimization strategies to minimize expensive information re-fetching

Use Cases

01Debugging large systems where the agent must remember historical file changes and decisions

02Managing long-running agent sessions that exceed standard context window limits

03Summarizing complex codebase exploration into actionable implementation plans

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add kalyanikhandare29/agent-skills-for-context-engineering context-compression

For use in Claude.ai and ChatGPT

Download Skill