litert-quantization-calib

Assists the user to calibrate, merge, and statically quantize litert LLM models (such as Gemma 3) in standard open-source (OSS) environments. Use when the user wants to run LLM calibration, merge task JSON results, align KV cache parameters across models, protect sensitive layers in Float32, or run quantized inference testing. Don't use for JAX/PyTorch custom quantization configurations or non-litert models.

Quality

82%

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

Quality

Content

65%

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 concrete, executable quantization commands and valuable domain-specific gotchas, but it is a monolithic single file with no progressive disclosure and lacks explicit validation/feedback checkpoints for its batch quantization workflow. Tightening the gotcha prose and adding verify-then-proceed steps would lift the weaker dimensions.

Suggestions

Add explicit validation checkpoints to the workflow (e.g. after calibration, verify the per-task JSON files exist under {output_save_dir}/{task_name} before merging; after quantization, run inference test step 4 and confirm non-garbage output before declaring success) to lift workflow_clarity.

Split the gotchas and command recipes into reference files (e.g. references/gotchas.md, references/commands.md) and keep SKILL.md as a concise overview that links to them, improving progressive_disclosure.

Tighten each gotcha to the rule plus a one-line rationale and remove the restated frontmatter intro to improve conciseness.

Dimension	Reasoning	Score
Conciseness	Mostly efficient and focused on LiteRT-specific gotchas Claude would not already know, but each gotcha carries a prose 'Gotcha' explanation plus a 'Rule' that could be tightened, and the opening paragraph restates the frontmatter. Matches the level-2 anchor of mostly efficient with some unnecessary explanation; not level 3 because padding remains.	2 / 3
Actionability	Provides fully executable, copy-paste-ready bash commands with concrete flags (e.g. the calibrate invocation with --use_profiler_based_calibration=True and --enable_min_max_calibration_update=True), matching the level-3 anchor of executable commands and specific examples. The single malformed 'bash cd ...' line is a minor blemish that does not undermine the overall copy-paste quality.	3 / 3
Workflow Clarity	The four core command patterns imply a clear sequence (calibrate -> merge -> quantize -> inference test), but there are no explicit validation checkpoints or feedback loops for these batch/destructive quantization operations; per the rubric, missing validation for batch operations caps workflow clarity at 2 rather than 3.	2 / 3
Progressive Disclosure	Content is well-sectioned with clear headers, but it is a single monolithic ~180-line file with no bundle files (references/scripts/assets absent) and no one-level-deep references signaled, so content that could be split (gotchas, command recipes) is inline. Matches the level-2 anchor of structure present but content that should be separate is inline; not level 1 because organization is genuinely clear.	2 / 3
	Total	9 / 12 Passed

Description

100%

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 specific, complete, and well-scoped: it names concrete actions, provides explicit Use/Don't-use triggers, and carves out a distinct niche. Voice is appropriately third person ('Assists the user'), so no specificity penalty applies.

Dimension	Reasoning	Score
Specificity	Lists multiple concrete actions — 'calibrate, merge, and statically quantize litert LLM models' plus 'align KV cache parameters', 'protect sensitive layers in Float32', and 'run quantized inference testing' — matching the level-3 anchor of several specific concrete actions.	3 / 3
Completeness	Explicitly answers both 'what' (calibrate, merge, statically quantize litert LLM models) and 'when' via an explicit 'Use when the user wants to...' clause plus a 'Don't use for...' exclusion, matching the level-3 anchor for both what AND when with explicit triggers.	3 / 3
Trigger Term Quality	Covers natural domain terms a user would actually say — 'LLM calibration', 'merge task JSON results', 'quantized inference testing', 'Gemma 3' — rather than generic jargon, matching the good-coverage anchor. Not the level below, which expects missing common variations; here the variations (calibrate/merge/quantize/test) are present.	3 / 3
Distinctiveness Conflict Risk	Targets a clear niche (litert LLM quantization) with distinct triggers and an explicit exclusion ('Don't use for JAX/PyTorch custom quantization configurations or non-litert models'), making conflict with other skills unlikely per the level-3 anchor.	3 / 3
	Total	12 / 12 Passed

Validation

100%

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 — 16 / 16 Passed

Validation for skill structure

No warnings or errors.

Repository: google-ai-edge/litert-torch
Path: litert_torch/generative/export_hf/experimental/calib/skills/tflite_quantization_calib/SKILL.md
Commit: 79bb9bb

Reviewed: 4 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.