playbooks

home / skills / eyadsibai / ltk / context-degradation

context-degradation skill

/plugins/ltk-core/skills/context-degradation

This skill helps diagnose context degradation in AI agents by identifying loss-in-middle, context poisoning, and distraction patterns to improve robustness.

npx playbooks add skill eyadsibai/ltk --skill context-degradation

Review the files below or copy the command above to add this skill to your agents.

Files (1)
SKILL.md
3.6 KB
---
name: context-degradation
description: Use when diagnosing agent failures, debugging lost-in-middle issues, understanding context poisoning, or asking about "context degradation", "lost in middle", "context poisoning", "attention patterns", "context clash", "agent performance drops"
version: 1.0.0
---

# Context Degradation Patterns

Language models exhibit predictable degradation as context grows. Understanding these patterns is essential for diagnosing failures and designing resilient systems.

## Degradation Patterns

| Pattern | Cause | Symptoms |
|---------|-------|----------|
| Lost-in-Middle | Attention mechanics | 10-40% lower recall for middle content |
| Context Poisoning | Errors compound | Tool misalignment, persistent hallucinations |
| Context Distraction | Irrelevant info | Uses wrong information for decisions |
| Context Confusion | Mixed tasks | Responses address wrong aspects |
| Context Clash | Conflicting info | Contradictory guidance derails reasoning |

## Lost-in-Middle

Information at beginning and end receives reliable attention. Middle content suffers dramatically reduced recall.

**Mitigation**:

```markdown
[CURRENT TASK]                      # At start (high attention)
- Goal: Generate quarterly report
- Deadline: End of week

[DETAILED CONTEXT]                  # Middle (less attention)
- 50 pages of data
- Supporting evidence

[KEY FINDINGS]                      # At end (high attention)
- Revenue up 15%
- Growth in Region A
```

## Context Poisoning

Once errors enter context, they compound through repeated reference.

**Entry pathways**:

1. Tool outputs with errors
2. Retrieved docs with incorrect info
3. Model-generated summaries with hallucinations

**Symptoms**:

- Tool calls with wrong parameters
- Strategies that take effort to undo
- Hallucinations that persist despite correction

**Recovery**:

- Truncate to before poisoning point
- Explicitly note poisoning and re-evaluate
- Restart with clean context

## Context Distraction

Even a single irrelevant document reduces performance. Models must attend to everything—they cannot "skip" irrelevant content.

**Mitigation**:

- Filter for relevance before loading
- Use namespacing for organization
- Access via tools instead of context

## Degradation Thresholds

| Model | Degradation Onset | Severe Degradation |
|-------|-------------------|-------------------|
| GPT-5.2 | ~64K tokens | ~200K tokens |
| Claude Opus 4.5 | ~100K tokens | ~180K tokens |
| Claude Sonnet 4.5 | ~80K tokens | ~150K tokens |
| Gemini 3 Pro | ~500K tokens | ~800K tokens |

## The Four-Bucket Approach

| Strategy | Purpose |
|----------|---------|
| **Write** | Save context outside window |
| **Select** | Pull relevant context in |
| **Compress** | Reduce tokens, preserve info |
| **Isolate** | Split across sub-agents |

## Counterintuitive Findings

1. **Shuffled haystacks outperform coherent** - Coherent context creates false associations
2. **Single distractors have outsized impact** - Step function, not proportional
3. **Needle-question similarity matters** - Dissimilar content degrades faster

## When Larger Contexts Hurt

- Performance degrades non-linearly after threshold
- Cost grows exponentially with context length
- Cognitive bottleneck remains regardless of size

## Best Practices

1. Monitor context length and performance correlation
2. Place critical information at beginning or end
3. Implement compaction triggers before degradation
4. Validate retrieved documents for accuracy
5. Use versioning to prevent outdated info clash
6. Segment tasks to prevent confusion
7. Design for graceful degradation
8. Test with progressively larger contexts

Overview

This skill diagnoses and mitigates context degradation in large language model agents. It focuses on patterns like lost-in-middle, context poisoning, distraction, confusion, and clash to help you debug agent failures and performance drops. Use it to triage attention-related failures and design resilient context management strategies.

How this skill works

The skill inspects agent context windows, attention distribution, and the sequence of inputs and tool outputs to locate where useful information is lost or corrupted. It identifies degradation patterns (e.g., lost-in-middle, poisoning, distraction) and recommends concrete remediation: truncation points, compaction, namespacing, or sub-agent isolation. It also correlates performance metrics with context length to detect thresholds where behavior shifts.

When to use it

  • When an agent forgets or misrecalls important details that were provided earlier.
  • When tool calls repeatedly use wrong parameters or hallucinations persist despite correction.
  • When adding more documents suddenly reduces accuracy or causes contradictory outputs.
  • When debugging multi-step workflows that drift off-target mid-way.
  • When evaluating system behavior as context length increases or when new docs are introduced.

Best practices

  • Place the most critical facts at the start or the end of the context window to maximize recall.
  • Filter and validate retrieved documents before loading to avoid context poisoning.
  • Implement compaction (summaries, embeddings) and set triggers to compress before degradation thresholds.
  • Use namespacing and clear delimiters to reduce cross-talk between unrelated content.
  • Segment complex tasks across sub-agents or stages to isolate reasoning and prevent confusion.
  • Monitor performance vs. context length to detect non-linear degradation and set safe limits.

Example use cases

  • Diagnosing a customer-support agent that starts contradicting earlier instructions after long dialogs.
  • Recovering from context poisoning caused by a buggy tool output that the agent kept reusing.
  • Designing a retrieval pipeline that selects and compresses only the most relevant passages.
  • Testing agent robustness by gradually increasing context size and observing attention shifts.
  • Rearchitecting a multi-step report generator to write key goals outside the long context body.

FAQ

How do I tell if my agent is 'lost in the middle'?

Compare recall or decision quality for items placed at the start, middle, and end of the context; a consistent drop for middle items indicates lost-in-middle.

What's the quickest recovery from context poisoning?

Truncate the session to before the poisoning point, explicitly mark the error, re-evaluate sources, and reload a clean, validated context.