How do I trigger this skill?

Trigger it when you notice degradation signs such as forgetting context, looping on fixes, generating inconsistent code, or user statements indicating that things are wrong or degraded.

What actions does it prescribe?

It prescribes stage-specific actions: re-reading relevant files, summarizing the plan, creating TASK.md and PROGRESS.md, delegating research, and in severe cases stopping coding and saving state.

How many stages are defined and what do they mean?

There are four stages: Mild, Moderate, Severe, and Critical. Each has defined signs and concrete actions to restore coherence and prevent drift.

Degradation Signs

Scanned

npx machina-cli add skill silvesterdivas/context-engineer/degradation-signs --openclaw

Files (1)

SKILL.md

2.9 KB

Context Degradation Self-Diagnosis

Context performance degrades in predictable ways as the window fills. Monitor for these signs and take corrective action.

Degradation Stages

Stage 1: Mild (60% context)

Signs:

Occasionally forgetting a detail from early in the conversation
Asking the user to re-confirm something already discussed
Slight inconsistencies in variable names or API shapes

Action:

Re-read the most critical file before making changes
Summarize the current plan before proceeding
Switch to more targeted tool use (Grep over Read)

Stage 2: Moderate (75% context)

Signs:

Forgetting which files you've already modified
Proposing changes that contradict earlier decisions
Missing imports or references that were established earlier
Generating code that doesn't match the project's patterns

Action:

Stop and create a brief status summary for yourself
Re-read only the files you're actively modifying
Delegate research tasks to the investigator agent
Consider creating TASK.md + PROGRESS.md

Stage 3: Severe (85% context)

Signs:

Hallucinating function signatures or file contents
Looping: trying the same fix repeatedly without progress
Producing code that won't compile/run due to misremembered APIs
Losing track of the overall goal

Action:

Stop coding immediately.
Create TASK.md and PROGRESS.md with current state
Tell the user: "I'm experiencing context degradation. Let me save our progress so we can continue in a fresh conversation."
Commit any working changes before wrapping up

Stage 4: Critical (90%+ context)

Signs:

Unable to maintain coherent multi-step reasoning
Contradicting yourself within a single response
Generating nonsensical or completely wrong code

Action:

Do not attempt any more code changes.
Save state immediately (TASK.md + PROGRESS.md)
End the conversation gracefully

Self-Check Protocol

When you suspect degradation, run this mental checklist:

Can I recall the project's main goal? If fuzzy → Stage 1+
Can I list the files I've modified? If unsure → Stage 2+
Can I describe the function I'm currently editing without re-reading it? If no → Stage 2+
Am I repeating a step I already tried? If yes → Stage 3
Does my code reference real APIs/functions? If unsure → Stage 3

Prevention

Use budget zones proactively (don't wait for degradation)
Delegate to subagents for any research that isn't immediately needed
Commit working changes frequently — smaller commits = safer state
Create fresh context files preemptively for complex tasks

Source

git clone https://github.com/silvesterdivas/context-engineer/blob/main/skills/degradation-signs/SKILL.mdView on GitHub

Overview

Context Degradation Self-Diagnosis helps detect when the conversation loses coherence, loops on fixes, or produces inconsistent code. It defines four stages (Mild to Critical) with explicit signs and corrective actions to maintain progress and accuracy.

How This Skill Works

The skill watches for degradation signals such as forgetting earlier details, contradicting changes, or hallucinated function signatures. When signals indicate a stage, it prescribes concrete steps: re-read critical files, summarize the plan, create TASK.md and PROGRESS.md, delegate research, or stop coding and save state. These steps are designed to prevent drift and help resume a fresh, accurate session.

When to Use It

Forgetting an early conversation detail and asking to reconfirm
Looping on the same fix without progress or repeating the same errors
Generating code that misaligns with project patterns or APIs
User comments indicating hallucination, wrongness, or context loss
Suspecting context degradation and needing a controlled pause with saved progress

Quick Start

Step 1: Monitor for signs like forgetting details, looping, or user distrust
Step 2: If detected, perform stage-appropriate actions (re-read, summarize, delegate, save state)
Step 3: Communicate progress or pause the session with saved TASK.md and PROGRESS.md

Best Practices

Monitor for explicit signs of degradation (forgetfulness, looping, inconsistencies, user warnings)
Switch to targeted tool use and re-read only the most relevant files when possible
Summarize the current plan before making changes
Delegate non-immediate research tasks to subagents
Create and save TASK.md and PROGRESS.md to capture state and next steps

Example Use Cases

Mild: You forget a detail from early in the conversation and re-read the critical file before proceeding
Moderate: You lose track of which files were modified and generate changes that contradict earlier decisions
Severe: You hallucinate function signatures and loop on the same fix without making progress
Critical: You cannot maintain coherent multi-step reasoning and begin producing nonsensical output
Preventive: You proactively create fresh context files and commit small, safe changes to prevent drift

Frequently Asked Questions

Add this skill to your agents