ov-gguf-debug-accuracy

Debug wrong or degraded output from a GGUF model running through the OpenVINO GGUF frontend or the ggml-openvino backend. Use when a GGUF model converts but produces garbage, repeated or drifting tokens, a cosine-similarity cliff between layers, output that diverges from llama.cpp CPU, or a broadcast/shape crash that appears only at decode or only for one architecture. Do NOT use for ops that fail to convert at all ("Translation for operation type ... is not implemented"), or for build and CMake failures.

Quality

86%

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

Quality

Content

86%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.

An efficient, well-sequenced debugging workflow with an explicit regression checkpoint and a single cleanly-signaled external reference; it is held just below top marks by deferring the executable technique to the referenced doc and leaving error-recovery feedback implicit.

Suggestions

Inline the bisection procedure's first concrete step (e.g. the exact comparison command against llama.cpp CPU) so the body is copy-paste executable rather than deferring all technique to the referenced doc.

Add an explicit error-recovery feedback loop after step 4 (e.g. 'If another architecture's output changes, bisect the shared-path change before re-running').

Name the specific debug env vars in the body instead of only listing them as contents of the referenced doc, so a user can act without opening it.

Dimension	Reasoning	Score
Conciseness	The four-step body is lean and assumes Claude's competence (e.g. 'If you are about to write out an op's math... generate it from ggml instead'); there is no padding or explanation of concepts Claude already knows, so every token earns its place.	5 / 5
Actionability	Steps give concrete directive guidance with specifics ('VIEW or op_case predicate', 'llama.cpp CPU', 'debug env vars') and a single one-level-deep doc pointer, but the core executable technique is deferred to the referenced file rather than inlined, so it is mostly- rather than fully-executable.	4 / 5
Workflow Clarity	The numbered sequence (read doc -> apply governing rule -> bisection -> post-fix regression) is clear with an explicit validation checkpoint in step 4 ('re-run the other supported architectures and confirm their classification and output are unchanged'), but an explicit error-recovery feedback loop is only implicit, capping it below 5.	4 / 5
Progressive Disclosure	The body is a concise overview that signals exactly one one-level-deep reference via a markdown link ([src/.../debugging_accuracy.md]); no bundle files are present and no nested references are introduced, matching the clear-navigation anchor-5 example.	5 / 5
	Total	18 / 20 Passed

Description

82%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.

A tight, well-targeted description with explicit what/when guidance and clear boundary exclusions; its only weakness is listing a single concrete action ('Debug') rather than several distinct operations.

Dimension	Reasoning	Score
Specificity	The description names the domain ('Debug wrong or degraded output from a GGUF model running through the OpenVINO GGUF frontend or the ggml-openvino backend') but offers essentially one action verb ('Debug'); the remaining clauses enumerate symptoms/scope rather than additional distinct actions, so it is not comprehensive enough for a 4.	3 / 5
Completeness	It explicitly answers both 'what' ('Debug wrong or degraded output...') and 'when' ('Use when a GGUF model converts but produces garbage... or a broadcast/shape crash...') with concrete trigger phrases and a negative boundary, matching the anchor-5 example.	5 / 5
Trigger Term Quality	It uses natural failure-mode phrases a developer would actually say ('garbage, repeated or drifting tokens', 'output that diverges from llama.cpp CPU', 'broadcast/shape crash') but omits the file extension (.gguf) and a few synonyms, falling short of comprehensive anchor-5 coverage.	4 / 5
Distinctiveness Conflict Risk	It carves a clear niche (OpenVINO GGUF frontend / ggml-openvino backend accuracy debugging) with distinct triggers and an explicit 'Do NOT use for...' exclusion, giving minimal conflict risk as in the anchor-5 example.	5 / 5
	Total	17 / 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
relative_links	Relative link issues: 1 suspicious	Warning

	Total	15 / 16 Passed

Repository: openvinotoolkit/openvino
Path: .claude/skills/ov-gguf-debug-accuracy/SKILL.md
Commit: f766feb

Reviewed: 3 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.