CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

ov-gguf

Develop and troubleshoot the OpenVINO GGUF frontend (src/frontends/gguf) and the llama.cpp ggml-openvino backend that uses it. Use when checking whether a .gguf or mmproj model is supported; when a .gguf file is rejected as an unsupported architecture; when enabling a llama.cpp model family (llama, qwen, gemma, phi, MoE, hybrid/recurrent) through arch_registry, DecoderBuilder, a ModelBuilder or an ArchitectureExtension; when adding a vision/audio projector or ProjectorExtension; when conversion fails with "Translation for operation type GGML_OP_* is not implemented" or an op translator/op_case in src/op/ needs work; when making a converted model stateful or GenAI/PagedAttention-ready; or when a GGUF model converts but gives garbage, drifting tokens, a llama.cpp mismatch or a decode-only shape failure; or when running or resuming GGUF frontend, backend and GenAI validation batches. Not for general llama.cpp usage or unrelated build/CI failures.

Quality

96%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Passed

No findings from the security scan

SKILL.md
Quality
Evals
Security

Quality

Content

93%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.

A dense, high-signal routing skill: the body holds only invariants, gotchas, and exact gates while pushing detail to verified one-level-deep references, with concrete environment variables, commands, and validation checkpoints throughout. The single gap is that error-recovery feedback loops are expressed as symptom-to-fix mappings rather than explicit gated sequences, leaving workflow clarity just below the top anchor.

DimensionReasoningScore

Conciseness

The body assumes full competence (no explanation of what GGUF, MoE, KV, SDPA, or PA are) and every sentence carries project-specific, non-derivable information such as "New translator or builder sources go into the explicit FRONTEND_SRCS list in tests/CMakeLists.txt, or the test binary fails to link" and "Do not tighten expect_near tolerances from an x86-only run". It matches the lean top anchor with no over-explanation to trim, rather than the level-4 anchor which expects minor instances of over-explanation.

5 / 5

Actionability

For an instruction/routing skill the guidance is fully executable: copy-paste-ready environment variables with exact values (OV_GGUF_Q4_K_ZP_F16=1, INFERENCE_PRECISION_HINT=f32, DYNAMIC_QUANTIZATION_GROUP_SIZE=0, OV_GGUF_REQUIRE_ARCH_FIXTURES=1), concrete invocations (gen_arch_fixtures.py --fetch, ov_gguf_frontend_tests unfiltered), exact test-suite names, and exact ownership boundaries ("the fix belongs in src/quant/, not a builder"). This matches the top anchor for copy-paste-ready commands covering the common cases, per the rubric note that instruction-only skills need concrete guidance rather than code.

5 / 5

Workflow Clarity

There is a clear entry step ("First identify the input path"), explicit ordering constraints (extensions before load(), converters before convert()), and real validation checkpoints for batch operations (run ov_gguf_frontend_tests unfiltered and check the exit status, "zero selected tests or suppressed required cases are missing coverage", verify required cases before a broad batch), so the batch-operation cap does not apply. It falls between the level-4 and level-5 anchors: symptom-to-fix mappings exist ("An unclaimed general.architecture needs a catalog row or ArchitectureExtension") but explicit validate-fix-revalidate feedback loops and per-task checklists are largely delegated to the referenced docs rather than stated as gated sequences in the body.

4 / 5

Progressive Disclosure

The body is a concise overview with a task-to-document routing table; all referenced bundle files exist (references/validation.md, wwb.md, preflight.md, build-runtime.md) and every linked anchor resolves (#family-contract, #multimodal-contracts, #scenario-environments, #test-selection-and-required-fixtures). References are one level deep, clearly signaled by task, and the body inlines only overview-level invariants, matching the top anchor for well-signaled, easily navigable structure.

5 / 5

Total

19

/

20

Passed

Description

100%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.

An exemplary description: third-person voice, a precise statement of what the skill covers, and an exhaustive set of natural trigger conditions including exact error messages, file extensions, and symptom phrasings, with an explicit negative boundary to prevent mis-triggering. Length is justified because every clause is a distinct trigger, not padding.

DimensionReasoningScore

Specificity

The description lists multiple concrete, comprehensive actions with named components and paths: "Develop and troubleshoot the OpenVINO GGUF frontend (src/frontends/gguf) and the llama.cpp ggml-openvino backend", "enabling a llama.cpp model family ... through arch_registry, DecoderBuilder, a ModelBuilder or an ArchitectureExtension", "adding a vision/audio projector or ProjectorExtension", "making a converted model stateful or GenAI/PagedAttention-ready", and "running or resuming GGUF ... validation batches". This matches the anchor for comprehensive concrete actions and exceeds the level-4 anchor (only minor gaps) because every capability of the skill is enumerated.

5 / 5

Completeness

It explicitly answers both questions: "what" is stated up front ("Develop and troubleshoot the OpenVINO GGUF frontend ... and the llama.cpp ggml-openvino backend that uses it") and "when" is given as a chain of explicit "Use when ..." triggers plus a negative boundary ("Not for general llama.cpp usage or unrelated build/CI failures"). This clearly matches the anchor requiring concrete trigger phrases for both what and when, not merely the level-4 anchor where the when-clause could be more specific.

5 / 5

Trigger Term Quality

Trigger coverage includes the file extension (.gguf, mmproj), symptom phrases users would naturally report ("converts but gives garbage, drifting tokens", "rejected as an unsupported architecture"), an exact error string ("Translation for operation type GGML_OP_* is not implemented"), and model-family names (llama, qwen, gemma, phi, MoE). This is comprehensive natural-term coverage including extensions and synonyms, matching the top anchor rather than the level-4 anchor that expects a few missing terms.

5 / 5

Distinctiveness Conflict Risk

The niche is tightly scoped to the OpenVINO GGUF frontend and its llama.cpp backend, with distinctly technical triggers (GGML_OP translators, ProjectorExtension, GenAI/PagedAttention) unlikely to appear in other skills, and the closing exclusion ("Not for general llama.cpp usage or unrelated build/CI failures") explicitly fences off the main overlap risks. Minimal conflict risk, matching the clear-niche top anchor.

5 / 5

Total

20

/

20

Passed

Validation

93%

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

CriteriaDescriptionResult

relative_links

Relative link issues: 13 suspicious

Warning

Total

15

/

16

Passed

Repository

openvinotoolkit/openvino

Path

.claude/skills/ov-gguf/SKILL.md

Reviewed

Table of Contents

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.