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

75%

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 ./skills/emerging-techniques/speculative-decoding/SKILL.md

Quality

Content

57%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.

The body is comprehensive and largely actionable with good method-selection guidance, but it is verbose with duplicated content across the body and reference files, lacks validation checkpoints, and ships a broken reference link.

Suggestions

Remove the duplicate conceptual/pseudocode content from Core Concepts that re-explains Quick Start, and trim the ASCII architecture and math-formulation blocks to reduce token cost (improves conciseness).

Fix the broken 'references/draft_model.md' link in See Also — either create the file or drop the reference — and signal references inline (e.g., 'See references/medusa.md for full training recipe') instead of inlining the same training code in the body (improves progressive_disclosure).

Add a validation/measurement checkpoint showing how to confirm the achieved speedup or token acceptance rate (e.g., a short benchmark snippet) so users can verify a configuration worked (improves workflow_clarity).

Dimension	Reasoning	Score
Conciseness	The ~450-line body duplicates material between Quick Start and Core Concepts (speculative_decode pseudocode, Medusa architecture, Lookahead class re-shown), and ASCII architecture diagrams plus mathematical formulations pad sections that could be tightened.	3 / 5
Actionability	Quick Start and the vLLM deployment block give mostly copy-paste-ready executable code with concrete hyperparameter values, but the first example uses torch.float16 without importing torch and Core Concepts mixes in non-executable pseudocode.	4 / 5
Workflow Clarity	Method-selection decision logic, the comparison table, and numbered algorithm steps provide a sequence, but there are no validation or measurement checkpoints (e.g., how to verify achieved speedup or acceptance rate).	3 / 5
Progressive Disclosure	Sections and a See Also list give structure, but references/draft_model.md is cited yet does not exist, training/architecture/hyperparameter content is duplicated verbatim between the body and medusa.md/lookahead.md rather than split out, and references are not signaled inline where the detail belongs.	3 / 5
	Total	13 / 20 Passed

Description

92%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 strong, well-structured description that clearly states capabilities and trigger conditions with concrete, niche-specific language. Its only weakness is slightly jargon-heavy trigger terms and a few missing common synonyms.

Dimension	Reasoning	Score
Specificity	Lists multiple specific concrete actions and techniques ('speculative decoding, Medusa multiple heads, and lookahead decoding', 'draft models, tree-based attention, Jacobi iteration, parallel token generation, and production deployment strategies'), giving comprehensive coverage of the domain.	5 / 5
Completeness	Explicitly answers both what ('Accelerate LLM inference using…') and when ('Use when optimizing inference speed…') with concrete trigger phrases, matching the anchor-5 example structure.	5 / 5
Trigger Term Quality	Includes natural trigger phrases ('optimizing inference speed', 'reducing latency for real-time applications', 'deploying models with limited compute') but leans heavily on technical jargon (Medusa, Jacobi iteration, tree-based attention) and omits common synonyms like 'faster' or 'speed up'.	4 / 5
Distinctiveness Conflict Risk	Targets a clear niche (speculative decoding / Medusa / lookahead decoding) with distinct, specific triggers and minimal overlap risk with general inference or training skills.	5 / 5
	Total	19 / 20 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: 4b35756

Reviewed: 1 day 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.