cu-lengths-attention-flow

Bilingual guide for understanding how cu_lengths controls attention behavior across ViT and LLM stages, and how patch_positions scope differs between the two

Quality

62%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Passed

No findings from the security scan

Fix and improve this skill with Tessl

tessl review fix ./.opencode/skills/cu-lengths-attention-flow/SKILL.md

Quality

Content

67%Weight 40%Scale 1-5

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

The content is a well-structured, accurate explainer with concrete code, file-to-role tables, and a clear mechanism progression that adds genuine repo-specific knowledge Claude would not already have. Its main weakness is token efficiency: the full English/Chinese duplication inflates size without adding reasoning value for a model that handles both languages.

Suggestions

Collapse the bilingual duplication: keep one primary language inline and move the other to a referenced file (e.g. README.zh.md), or use a single concise bilingual glossary instead of mirroring every paragraph.

Consider moving the SFT attention-mask cu_seqlens derivation and the sequence-parallelism padding detail into a separate references file referenced from the main body to shorten the core SKILL.md.

Add a short "How to use this skill" line stating the recommended reasoning sequence (locate cu_lengths shape → determine packed vs non-packed → trace attention domain) to make the workflow explicit rather than implicit in the section order.

Dimension	Reasoning	Score
Conciseness	The body is high-signal and avoids explaining concepts Claude already knows, but every section is duplicated in English and Chinese, roughly doubling the token count; that deliberate bilingual padding keeps it at mostly-efficient-but-could-be-tightened rather than lean.	3 / 5
Actionability	Concrete real code excerpts (the cu_lengths construction, the forward-branch condition, the PackedSeqParams call), exact file paths, and shape conditions give mostly-executable guidance, though the snippets are illustrative extracts rather than complete copy-paste scripts.	4 / 5
Workflow Clarity	The mechanism is laid out as a clearly numbered, logically sequenced progression (generation → forward branching → LLM behavior → SP padding) with tables and a diagram; as a non-destructive explainer it needs no validation checkpoints, so it sits just below the explicit-checklist anchor.	4 / 5
Progressive Disclosure	The single SKILL.md is well-organized with clear headers, tables, and a diagram and needs no external references (none exist in the bundle), but it is a monolithic ~278-line file where the SFT path and SP-padding detail could plausibly live in separate referenced files, leaving minor organization gaps.	4 / 5
	Total	15 / 20 Passed

Description

57%Weight 40%Scale 1-5

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

The description is highly specific and clearly distinguishable, naming the exact mechanism and model it covers, but it omits an explicit "when to use" trigger clause and relies on technical jargon without natural-language synonyms. It is a strong what-statement that lacks the matching when-statement.

Suggestions

Add an explicit trigger clause, e.g. "Use when debugging packed vs non-packed attention, reasoning about cross-sample isolation, or tracing why patch_positions grouping does not carry into the LLM in LLaVA-OneVision2."

Soften the jargon-only triggers by including a natural-language phrasing a user might actually say, such as "attention boundaries" or "packed sequence isolation".

Reframe "guide for understanding" around concrete actions Claude takes (e.g. "Explains and traces how cu_lengths and patch_positions control attention...") to lift specificity above 3.

Dimension	Reasoning	Score
Specificity	Names a concrete domain ("how cu_lengths controls attention behavior across ViT and LLM stages" and "patch_positions scope differs") but frames it as a guide "for understanding" rather than listing multiple concrete actions, so it sits at the names-domain-with-1-2-concrete-anchors level rather than a comprehensive action list.	3 / 5
Completeness	The "what" is clear (a bilingual guide explaining the two attention-control mechanisms), but there is no "Use when..." clause or equivalent explicit trigger guidance, which caps completeness at 3 per the judging guidelines.	3 / 5
Trigger Term Quality	Relevant identifiers (cu_lengths, patch_positions, ViT/LLM stages, attention behavior) appear but are technical jargon with no natural-language synonyms or phrasings a user would vary, matching the some-relevant-keywords-but-missing-variations anchor.	3 / 5
Distinctiveness Conflict Risk	The description targets a uniquely narrow mechanism in a specific model (LLaVA-OneVision2 cu_lengths/patch_positions attention scoping), giving it a clear niche with minimal overlap risk against any other skill.	5 / 5
	Total	14 / 20 Passed

Validation

93%

Warnings & errors only

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 15 / 16 Passed

Validation for skill structure

Criteria	Description	Result
metadata_version	'metadata.version' is missing	Warning

	Total	15 / 16 Passed

Repository: EvolvingLMMs-Lab/LLaVA-OneVision-2
Path: .opencode/skills/cu-lengths-attention-flow/SKILL.md
Commit: f1b6702

Reviewed: 2 days ago

Table of Contents

Discovery Implementation Validation

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.