blip-2-vision-language

Vision-language pre-training framework bridging frozen image encoders and LLMs. Use when you need image captioning, visual question answering, image-text retrieval, or multimodal chat with state-of-the-art zero-shot performance.

Quality

78%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Low

Low-risk findings worth noting

Fix and improve this skill with Tessl

tessl review fix ./18-multimodal/blip-2/SKILL.md

Quality

Content

72%

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

The body is highly actionable with extensive executable code and a sound progressive-disclosure structure that offloads advanced material to two well-linked reference files. Its weaknesses are verbosity (redundant Transformers+LAVIS coverage and full application classes) and the absence of validation checkpoints in the batch workflows.

Suggestions

Trim redundancy by leading with one primary library (Transformers) and folding the LAVIS alternative into the advanced-usage reference.

Add explicit validation/verification steps to the batch and retrieval workflows (e.g. checking decoded output, sanity-checking similarity scores) to introduce feedback loops.

Consider moving the full ImageCaptioner/VisualQA/ImageSearchEngine class examples into a reference file, keeping the body to concise illustrative snippets.

Dimension	Reasoning	Score
Conciseness	The ~555-line body shows the same tasks twice (Transformers and LAVIS), includes three full application-style classes and ASCII architecture art; mostly relevant but could be tightened significantly, matching the mostly-efficient-but-could-be-tighter anchor; not 1 because it avoids explaining basic concepts Claude already knows, not 3 because it is far from lean.	2 / 3
Actionability	Abundant executable, copy-paste-ready code with concrete model identifiers and parameters (e.g. Blip2ForConditionalGeneration with torch_dtype, device_map, quantization configs); not 2 because the examples are complete rather than pseudocode.	3 / 3
Workflow Clarity	Workflows are presented as runnable classes with no explicit validation checkpoints, and batch operations (caption_batch, batch processing) lack verification steps, which caps this dimension at 2 per the batch-operation guideline; not 3 because no validate->fix->retry feedback loops are shown.	2 / 3
Progressive Disclosure	Two real, well-named, one-level-deep references (references/advanced-usage.md, references/troubleshooting.md) are clearly signaled via markdown links and hold the genuinely advanced/deployment content, matching the clear-overview-with-well-signaled-references anchor; not 2 because the references are clearly signaled and appropriately split rather than left vague.	3 / 3
	Total	10 / 12 Passed

Description

85%

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 strong: it states what the skill does and gives an explicit "Use when" trigger covering several concrete VL tasks. The main weakness is that several trigger terms are somewhat technical and miss everyday phrasings users would actually say; the trailing "state-of-the-art zero-shot performance" is a mild over-claim.

Suggestions

Add common conversational trigger terms (e.g. "describe an image", "what's in this picture", "image search") alongside the technical ones to improve natural-keyword coverage.

Drop or soften the "state-of-the-art zero-shot performance" over-claim to keep the description factual and buzzword-free.

Dimension	Reasoning	Score
Specificity	Lists multiple concrete tasks — "image captioning, visual question answering, image-text retrieval, or multimodal chat" — matching the multiple-specific-actions anchor; not level 2 because the action coverage is comprehensive rather than partial.	3 / 3
Completeness	Explicitly answers both what ("Vision-language pre-training framework bridging frozen image encoders and LLMs") and when via an explicit "Use when..." trigger clause; not 2 because the when is explicit rather than implied.	3 / 3
Trigger Term Quality	Relevant keywords are present but lean technical ("visual question answering", "image-text retrieval") and omit common conversational variations a user would naturally say (e.g. "describe an image", "image search"), matching the some-keywords-missing-variations anchor; not 3 due to those gaps, not 1 because the terms are user-facing rather than pure jargon.	2 / 3
Distinctiveness Conflict Risk	A clear vision-language niche (frozen-encoder + LLM bridging) with distinct task triggers makes wrongful-skill triggering unlikely; not 2 because the framing is more specific than a generic document/file scope.	3 / 3
	Total	11 / 12 Passed

Validation

81%

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 — 13 / 16 Passed

Validation for skill structure

Criteria	Description	Result
skill_md_line_count	SKILL.md is long (565 lines); consider splitting into references/ and linking	Warning
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning
relative_links	Relative link issues: 7 missing	Warning

	Total	13 / 16 Passed

Repository: Orchestra-Research/AI-Research-SKILLs
Path: 18-multimodal/blip-2/SKILL.md
Commit: 773a529

Reviewed: about 12 hours 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.