speculative-decoding

Accelerate LLM inference using speculative decoding, Medusa multiple heads, and lookahead decoding techniques. Use when optimizing inference speed (1.5-3.6× speedup), reducing latency for real-time applications, or deploying models with limited compute. Covers draft models, tree-based attention, Jacobi iteration, parallel token generation, and production deployment strategies.

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 content is highly actionable with real executable code and specific tuning guidance, but it over-explains concepts Claude already knows, lacks validation checkpoints in its deployment workflow, and ships a broken reference path plus inline content that overlaps the existing bundle files.

Suggestions

Add a verification checkpoint after enabling speculative decoding (e.g., measure tokens/sec and compare output quality against the non-speculative baseline) so the deployment workflow has an explicit feedback loop.

Trim the conceptual prose ('Idea', 'Innovation', 'Mathematical formulation') that restates how speculative decoding, Medusa, and Jacobi iteration work, since Claude already knows these techniques.

Fix the 'See Also' navigation: either create references/draft_model.md or point that line to an existing file, and move the duplicated Medusa/Lookahead core-concept and training detail into the existing reference files to keep SKILL.md a lean overview.

Dimension	Reasoning	Score
Conciseness	The body is ~450 lines and re-explains concepts Claude already knows ('Idea: Use small draft model...', the Jacobi 'Mathematical formulation', 'Innovation: Add multiple prediction heads'), so it is mostly efficient with concrete code but padded with conceptual prose that could be trimmed.	2 / 3
Actionability	Provides concrete, copy-paste-ready examples using real APIs (transformers assisted generation, Medusa medusa_generate, vLLM speculative_model config) plus specific hyperparameter values (K=4, posterior_threshold=0.09, W=15, N=5) and a decision function.	3 / 3
Workflow Clarity	There is logical sequencing (Installation -> Quick Start -> Best Practices decision tree -> Production Deployment) but no validation checkpoints: nothing tells Claude to verify the speedup, measure acceptance rate, or confirm output quality is unchanged after enabling speculative decoding.	2 / 3
Progressive Disclosure	References are one level deep and clearly signaled, but 'See Also' points to references/draft_model.md which does not exist in the bundle (only medusa.md and lookahead.md are present), and substantial core-concept/training content that could live in the reference files is duplicated inline.	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.

A strong description that states concrete capabilities, names the specific techniques, and includes an explicit 'Use when...' trigger clause covering natural scenarios. It uses third-person voice and avoids vague fluff or over-claims.

Dimension	Reasoning	Score
Specificity	Lists multiple concrete actions ('Accelerate LLM inference', 'optimizing inference speed', 'reducing latency', 'deploying models with limited compute') plus a specific technique catalog (speculative decoding, Medusa, lookahead, tree-based attention, Jacobi iteration), matching the 'lists multiple specific concrete actions' anchor.	3 / 3
Completeness	Clearly answers both 'what' ('Accelerate LLM inference using speculative decoding, Medusa multiple heads, and lookahead decoding techniques') and 'when' with an explicit 'Use when...' trigger clause.	3 / 3
Trigger Term Quality	The 'Use when optimizing inference speed... reducing latency for real-time applications, or deploying models with limited compute' clause uses natural terms a user would say, and names the specific techniques someone seeking this skill would mention.	3 / 3
Distinctiveness Conflict Risk	Named techniques (speculative decoding, Medusa, lookahead decoding) carve a clear inference-acceleration niche with distinct triggers; the only slightly broad phrasing ('limited compute') is anchored by the specific methods.	3 / 3
	Total	12 / 12 Passed

Validation

87%

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 — 14 / 16 Passed

Validation for skill structure

Criteria	Description	Result
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning
referenced_paths_exist	Referenced path issues: 1 missing	Warning

	Total	14 / 16 Passed

Repository: OpenLAIR/dr-claw
Path: skills/emerging-techniques/speculative-decoding/SKILL.md
Commit: 6da7f7c

Reviewed: about 6 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.