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

67%

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 ./backend/cli/skills/ml-inference/speculative-decoding/SKILL.md

Quality

Content

50%

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

The body is information-rich and largely actionable, but it is verbose for a SKILL.md, mixes executable code with unjustified pseudocode, lacks validation checkpoints, and partially defeats its own progressive disclosure by duplicating reference content inline and pointing to a missing draft_model.md reference.

Suggestions

Move the Medusa architecture and Lookahead math deep dives out of SKILL.md into references/medusa.md and references/lookahead.md, leaving a concise overview with one-level-deep links, and either create the missing references/draft_model.md or remove that 'See Also' entry.

Replace the illustrative pseudocode blocks (speculative_decode, LookaheadDecoding class, select_draft_model, the Best Practices decision block) with either executable code or explicitly labeled pseudocode with a justification, since the rubric penalizes unjustified pseudocode.

Add a short validation/verification step to the deployment and method-selection workflows (e.g., verify acceptance rate / speedup after switching methods) so the workflow includes an explicit checkpoint.

Dimension	Reasoning	Score
Conciseness	The ~450-line body contains genuinely useful code but duplicates material that belongs in references (Medusa architecture and lookahead math are re-explained inline despite references/medusa.md and lookahead.md existing) and restates concepts Claude already knows; it is not score 3 because notable content could be tightened or offloaded, and not score 1 because the bulk is specific rather than padded beginner prose.	2 / 3
Actionability	Strong executable examples exist (transformers assisted generation, Medusa medusa_generate, vLLM speculative_model, a real PyTorch training loop), but several Core Concepts and Best Practices blocks are non-executable pseudocode presented without justification (speculative_decode(), the LookaheadDecoding class with undefined methods, select_draft_model(), the decision pseudocode), fitting the 'pseudocode instead of executable code' anchor; it is not score 3 because of these unjustified pseudocode blocks, and not score 1 because substantial real, copy-paste-ready code is present.	2 / 3
Workflow Clarity	A logical sequence exists (Installation -> Quick Start -> Core Concepts -> Best Practices) plus a method-selection decision guide, but there are no validation checkpoints or feedback loops and the decision logic is pseudocode; it is not score 3 because explicit validation/error-recovery steps are absent, and not score 1 because the structural progression is clear.	2 / 3
Progressive Disclosure	References are signaled via a 'See Also' section, but the body duplicates reference-suitable content inline (Medusa/lookahead deep dives) and references/medusa.md and lookahead.md exist while the referenced references/draft_model.md is missing; this matches the 'content that should be separate is inline' anchor. It is not score 1 because organization is reasonable and references are only one level deep, and not score 3 because of the inline duplication and the broken draft_model.md path.	2 / 3
	Total	8 / 12 Passed

Description

85%

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, third-person description that clearly answers both what and when with an explicit 'Use when' clause and a well-defined niche. Its main weakness is trigger-term quality: the keyword mix is jargon-heavy and misses common user phrasings such as 'speed up' or 'throughput'.

Suggestions

Add common natural trigger phrasings to the 'Use when' clause such as 'speeding up inference', 'faster generation', or 'improving throughput' alongside the existing terms.

Move deep jargon ('Jacobi iteration', 'tree-based attention') out of the trigger clause and into the capability list so the trigger reads as terms a user would naturally say.

Consider noting a complementary negative boundary (e.g., 'Not for training-time or memory optimization') to further reduce conflict risk.

Dimension	Reasoning	Score
Specificity	Names concrete capabilities and multiple specific techniques ('speculative decoding, Medusa multiple heads, and lookahead decoding techniques' plus 'draft models, tree-based attention, Jacobi iteration, parallel token generation'), matching the multiple-specific-actions anchor; it is not score 2 because the action set is comprehensive rather than partial.	3 / 3
Completeness	It explicitly states what the skill does ('Accelerate LLM inference using...') and when to use it via an explicit 'Use when...' clause with concrete triggers, satisfying both halves; it is not score 2 because the 'when' is explicit rather than merely implied.	3 / 3
Trigger Term Quality	The 'Use when' clause includes relevant natural triggers ('optimizing inference speed', 'reducing latency', 'deploying models with limited compute') but is dense with jargon ('Jacobi iteration', 'tree-based attention') and omits common phrasings users would say like 'speed up', 'faster inference', or 'throughput', so it does not reach the full-coverage anchor.	2 / 3
Distinctiveness Conflict Risk	It carves a clear niche (LLM inference acceleration via speculative-decoding family techniques) with distinct triggers unlikely to fire for unrelated skills; it is not score 2 because the scope is narrowly specialized rather than broadly overlapping.	3 / 3
	Total	11 / 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: synthetic-sciences/openscience
Path: backend/cli/skills/ml-inference/speculative-decoding/SKILL.md
Commit: 52845c3

Reviewed: about 12 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.