context-degradation-review
Skill16 starsChanged 5 months agoMITView on GitHub
Review agent work for context poisoning, lost-in-the-middle failures, distraction, context clash, and stale carryover. Use before long-session handoffs, SET runner handoffs, skill changes, memory writes, or any task where too much or conflicting context may degrade behavior.
In this repository94 other files it gives its agents.
- agent-friction-ledgerSkill
- agent-learning-layer-triageSkill
- agent-operations-contractSkill
- agents-best-practicesSkill
- agent-tool-contract-reviewSkill
- anti-slop-reviewSkill
- architecture-deepening-reviewSkill
- assumption-excavationSkill
- authorized-security-routerSkill
- book-to-skillSkill
- bounded-evaluationSkill
- bounded-growth-loopSkill
- bounded-orchestration-contractSkill
- brief-first-executionSkill
- browser-verificationSkill
- bug-evidence-protocolSkill
- codex-review-loopSkill
- compaction-survivalSkill
- complexity-optimizerSkill
- confidence-fragility-reviewSkill
- corporate-surfaceSkill
- delivery-baseline-auditSkill
- delivery-preflight-gateSkill
- design-critique-polishSkill
- designmd-brand-kitSkill
- design-register-bootstrapSkill
- diagnoseSkill
- doc-grounded-grillingSkill
- doc-to-lora-evaluatorSkill
- durable-context-maintenanceSkill
- dynamic-workflow-packetsSkill
- editorial-surfaceSkill
- evidence-ledger-researchSkill
- filesystem-context-disciplineSkill
- fluid-interaction-reviewSkill
- frontend-product-builderSkill
- frontend-taste-layerSkill
- git-native-context-contractSkill
- goal-loop-designerSkill
- graph-guided-code-readingSkill
- handoffSkill
- happy-handoffSkill
- html-brief-artifactSkill
- html-diagram-artifactSkill
- hypothesis-diversificationSkill
- idea-to-ship-gatesSkill
- knowledge-base-enrichmentSkill
- lean-context-layoutSkill
- local-inference-tuningSkill
- loop-hardening-contractSkill
- loopops-protocolSkill
- loop-readiness-reviewSkill
- lottie-motion-builderSkill
- minimal-diff-builderSkill
- motion-review-gateSkill
- overengineering-reviewSkill
- personal-workspace-routerSkill
- phase-spec-executionSkill
- pipeline-readiness-gateSkill
- plan-to-issuesSkill
- The first 60 shown.
The file
We split this skill into the rules it gives an agent
8 rules. About 0 are common, 8 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈8
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 12
Use this skill to find when the context itself is making the agent worse.
- Distinctive: line 35
Check for the five failure modes above.
- Distinctive: line 37
remove or ignore a stale source.
- Distinctive: line 40
ask for clarification only when the conflict blocks action.
- Distinctive: line 57
Do not use this as a generic writing critique.
- Distinctive: line 58
Do not remove context just because it is long.
- Distinctive: line 59
Do not treat memory as higher authority than the current user message.
- Distinctive: line 60
Do not let stale rollout context override an explicit stop, pause, or changed task.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Context Degradation Review
Use this skill to find when the context itself is making the agent worse.
Use When
- a long session, handoff, or compaction summary may contain stale or conflicting instructions;
- a SET bundle, AGENTS.md, SKILL.md, or memory write mixes useful facts with outdated task history;
- the agent keeps optimizing for irrelevant prior work;
- tool output, logs, or retrieved docs may bury the decisive evidence;
- multiple sources disagree and the answer treats them as one coherent contract.
Failure Modes
lost-in-middle: key constraints are present but buried between low-signal material.context-poisoning: incorrect, stale, or speculative text is treated as authority.distraction: large but irrelevant context pulls work away from the current task.context-clash: two instructions, docs, memories, or artifacts conflict.stale-carryover: a prior goal continues after the user changed or stopped the task.
Workflow
- Identify the active task, current authority order, and expected output.
- List the context sources used: user message, repo files, memory, docs, tool output, web sources, generated summaries.
- Mark each source as
authoritative,supporting,stale-risk, orirrelevant. - Check for the five failure modes above.
- Propose the smallest context repair:
- remove or ignore a stale source;
- promote the decisive constraint near the active plan;
- split task history from durable guidance;
- ask for clarification only when the conflict blocks action;
- rewrite a handoff or memory proposal with source boundaries.
- Continue only after the active context is narrow enough to act safely.
Output Shape
Use:
- verdict:
CONTEXT_OK,CONTEXT_DEGRADED, orCONTEXT_BLOCKED; - active task;
- context sources reviewed;
- degradation findings by mode;
- smallest repair;
- remaining risk.
Guardrails
- Do not use this as a generic writing critique.
- Do not remove context just because it is long; remove or demote it only when it is low-signal, stale, or conflicting.
- Do not treat memory as higher authority than the current user message.
- Do not let stale rollout context override an explicit stop, pause, or changed task.
Provenance
Adapted from context degradation patterns in muratcankoylan/Agent-Skills-for-Context-Engineering, narrowed into an ABVX review lens.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

