How to use the Context Degradation Diagnostics agent skill

What is an agent skill and what does this one do?

An agent skill is a folder of instructions and optional scripts that an AI agent can load when a task matches the skill's description. This one is named context-degradation and is built around a SKILL.md file that describes when to activate and how to reason about context failures.

The skill names five failure patterns: lost-in-middle, poisoning, distraction, confusion, and clash. For each, it lists detection signals and mitigation strategies. Examples of the suggested fixes include placing critical facts at the beginning or end of context, removing poisoned content rather than correcting over it, separating unrelated tasks into different context windows, and marking contradictions with explicit source precedence.

• Lost-in-middle: information in the middle of a long context is recalled less reliably; move key facts to the start or end.
• Poisoning: a wrong fact, tool error, or hallucination compounds; truncate to before it or restart with verified context.
• Distraction: irrelevant documents degrade results; filter before loading and fetch on demand.
• Confusion: constraints from one task leak into another; isolate tasks.
• Clash: correct but contradictory sources; set priority rules and filter outdated versions.

Install the skill from the ZIP or the Claude apps

Download the ZIP for this skill from this page. Then choose one of the following options.

For Claude Code, unzip the download into your user skills folder ~/.claude/skills/, or into a project's .claude/skills/ folder if you want the skill limited to one project.

In the Claude apps, upload the ZIP under Settings → Capabilities → Skills.

• Personal use across projects: unzip into ~/.claude/skills/.
• Single project: unzip into the project's .claude/skills/ folder.
• Claude apps: upload the ZIP under Settings → Capabilities → Skills.

Check what the skill needs before you run anything

The skill's included Python script requires Python 3. It uses only standard library modules such as random and re, so no packages need to be installed for it to run. The skill does run scripts, and it does not use the network.

The skill's guidance is plain text you can read in SKILL.md and references/patterns.md before you rely on it.

• Requires: Python 3 for scripts/degradation_detector.py.
• Runs scripts: yes.
• Uses network: no.
• Files: SKILL.md, references/patterns.md, scripts/degradation_detector.py.

Trigger the skill with natural requests

You do not need a special command. The skill is meant to activate when your request matches its description, such as debugging an agent that gives incorrect or irrelevant output, or investigating lost-in-middle behaviour. Describe the symptom in your own words and say what you want the agent to do with it.

Good requests name the symptom, the context involved, and the outcome you want.

• "My agent ignores the policy I put at the top of a 30-page prompt. Diagnose why and suggest placement changes."
• "After 40 turns, the agent started mixing up two API versions. Check whether this is context clash and propose a precedence rule."
• "Our retrieval step adds ten documents per query. Review whether distractors could explain the wrong answers."
• "The agent calls tools meant for the billing task while answering a support question. Look for context confusion."

Run the demonstration script

The bundled scripts/degradation_detector.py simulates degradation metrics for demonstration. Its attention numbers are simulated, not read from a real model, so use it to see how the output is structured and to discuss the approach, not as evidence about your system.

Run it with Python 3 from the skill's scripts folder, for example by asking the agent to execute it, or run python3 scripts/degradation_detector.py from the skill folder. The source does not document the script's options, so check its contents before relying on any output.

• Treat the script's numbers as illustrative only.
• Use the references/patterns.md file for the measurement approach it describes.

Use the patterns on your own system

Start by reproducing the symptom with a fixed test input. Then check the signals for each pattern: correct information that the model ignores suggests lost-in-middle; a single wrong fact that keeps reappearing suggests poisoning; a mix of constraints from two tasks suggests confusion.

Apply the matching fix and test again. Recovery from poisoning means removing the bad content, not adding a correction on top of it.

• Reproduce the failure with the same input before changing anything.
• Match the symptom to one of the five patterns.
• Change one thing at a time and re-run the test.
• Do not treat simulated script metrics as real attention measurements.