Conciseness

The ~220-line body re-states the frontmatter description verbatim in the intro, extensively explains concepts Claude already knows (the lost-in-middle effect, attention sinks, "Models do not have a mechanism to 'skip' irrelevant context"), and inlines time-sensitive version data (a GPT-5.2/Claude Opus 4.5/Gemini 3 threshold table, "80.9% on SWE-bench Verified", "81.2% MMMU-Pro") that the guidelines say should penalize conciseness outside a deprecated section. This matches 'noticeably verbose; several unnecessary explanations or padded sections' — above score 1 because the material is domain-specific rather than generic library tutorial, below score 3 because the padding and redundancy are pervasive, not occasional.

Actionability

There is real concrete guidance — the Write/Select/Compress/Isolate four-bucket approach, "Use observation masking to replace verbose tool outputs with compact references", "Apply relevance filtering before loading retrieved documents", and the 8-item Guidelines checklist — but it stays at the strategy level: no executable code or commands, compaction triggers are named without thresholds, and the two Examples blocks are illustrative layouts rather than runnable artifacts. This fits 'some concrete guidance but incomplete' — below score 4 because a practitioner cannot execute most of the advice without filling in the specifics themselves.

Workflow Clarity

A rough diagnostic flow exists (symptoms like "degraded output quality on tasks that previously succeeded, tool misalignment" → "consider context poisoning" → recovery options "truncating context to before the poisoning point... or restarting with clean context"), but it is never presented as an explicit sequence, and there are no validation checkpoints confirming a diagnosis or a mitigation's effect. This fits 'sequence present but checkpoints missing or implicit' — below score 4 because no ordered multi-step procedure is laid out anywhere, above score 2 because detection and recovery steps are at least articulated per-pattern.

Progressive Disclosure

Section headers provide decent structure, but no bundle files exist (no references/, scripts/, or assets/), and the one internal reference — "Degradation Patterns Reference - Detailed technical reference" — carries no path or link and does not resolve to any file. Material that clearly belongs in a reference file (the model threshold table, per-pattern deep dives) is inlined in a ~220-line body. This matches 'some structure... references present but not clearly signaled; content that should be separate is inline' — below score 4 because the reference is dangling and detail is inlined, above score 2 because the body is genuinely well-sectioned and navigable.

Specificity

The description names the domain ("predictable degradation patterns as context length increases") but states no concrete skill actions — "diagnosing failures and designing resilient systems" are generic user goals, not capabilities the skill performs. It never says what the skill actually does (e.g., identify patterns, apply mitigation strategies), matching the anchor 'names the domain but actions are minimal or generic'. It is above score 1 because the domain is specific and clearly named, and below score 3 because not even 1-2 concrete actions are listed.

Completeness

The 'what' is vague — the text describes the phenomenon ("Language models exhibit predictable degradation patterns") rather than what the skill does, and the 'when' is entirely absent with no 'Use when...' clause or equivalent trigger guidance. This matches the anchor 'has a vague what and no when', and the rubric guideline that a missing 'Use when' clause caps completeness applies with full force here.

Trigger Term Quality

Only "context length" and "degradation" function as keywords; everything else is academic phrasing ("predictable degradation patterns", "designing resilient systems"). Natural phrases a user would say — "long context", "lost in the middle", "context window limits", "agent getting worse over long conversations" — are absent, matching the anchor 'one or two generic keywords; missing the natural phrases users say'. Below score 3 because the phrasing is conceptual rather than user-utterable, and there is no 'Use when' framing to surface variations.

Distinctiveness Conflict Risk

"Context degradation" carves out a recognizable niche, but the framing "diagnosing failures and designing resilient systems" overlaps directly with sibling skills the body itself lists (context-optimization for mitigation, evaluation for detection/measurement). It fits 'somewhat specific but could still overlap with similar skills' — below score 4 because the description's stated purposes are shared with closely related skills in the same collection.