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

72%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Passed

No findings from the security scan

Fix and improve this skill with Tessl

tessl review fix ./backend/cli/skills/ml-inference/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.

A thorough, code-rich reference with strong actionability and a clear method comparison, but it is verbose for a SKILL.md and inlines substantial detail that belongs in the existing reference files. Tightening the body and routing depth into references/medusa.md and references/lookahead.md would lift the conciseness and progressive-disclosure scores.

Suggestions

Move the Core Concepts explanations, the Medusa training loop, and the hybrid/advanced patterns into references/medusa.md and references/lookahead.md, keeping SKILL.md as a concise overview with inline links at each relevant section instead of a 'See Also' footer.

Trim narrative re-explanations of well-known concepts (the speculative-decoding 'Idea'/'Algorithm' prose, the Jacobi math formulation, ASCII tree art) in favor of a one-line summary plus a runnable snippet.

Make the Core Concepts code examples fully runnable by defining or importing the helper functions they call (softmax, generate_ngram, verify), or label them explicitly as illustrative pseudocode so the actionability gaps are intentional.

Dimension	Reasoning	Score
Conciseness	Mostly efficient with extensive concrete code, but it re-explains concepts Claude largely already knows (the speculative-decoding 'Idea' and 'Algorithm' narrative, the Jacobi mathematical formulation, ASCII tree diagrams), adding length that could be trimmed without losing actionable value.	3 / 5
Actionability	Provides mostly executable, copy-paste-ready guidance across transformers assisted generation, Medusa, lookahead, training Medusa heads, and vLLM deployment, with only minor gaps — the Core Concepts 'speculative_decode' function and 'LookaheadDecoding' class use undefined helpers (softmax, generate_ngram, verify), reading as illustrative pseudocode rather than fully runnable code.	4 / 5
Workflow Clarity	A logical section sequence (When to Use → Installation → Quick Start → Core Concepts → Best Practices) and a method-selection decision block exist, but this is a technique survey rather than a sequenced process, and there are no explicit validation or feedback checkpoints; no destructive/batch cap applies, so it stays at 3 rather than lower.	3 / 5
Progressive Disclosure	The two reference files (references/medusa.md, references/lookahead.md) exist and are one level deep, but they are signaled only in a 'See Also' footer rather than inline at the relevant sections, and a large amount of detailed material (Core Concepts explanations, training Medusa heads, hybrid patterns) is inlined in SKILL.md that could live in those reference files.	3 / 5
	Total	13 / 20 Passed

Description

87%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 concretely answers both what the skill does and when to use it, with explicit trigger phrases and a distinct niche. Slight room to tighten the 'Covers...' list into true actions and add more colloquial trigger synonyms.

Dimension	Reasoning	Score
Specificity	Lists several concrete actions ('Accelerate LLM inference', 'reducing latency', 'deploying models with limited compute') and enumerates covered techniques (draft models, tree-based attention, Jacobi iteration, parallel token generation, production deployment), but several of the 'Covers...' items are topics rather than actions, leaving minor gaps versus a fully comprehensive action list.	4 / 5
Completeness	Clearly states what it does ('Accelerate LLM inference using speculative decoding, Medusa multiple heads, and lookahead decoding techniques') and explicitly states when to use it with a 'Use when' clause and concrete triggers ('optimizing inference speed (1.5-3.6× speedup), reducing latency for real-time applications, or deploying models with limited compute').	5 / 5
Trigger Term Quality	Good coverage of terms a user optimizing inference would say ('inference speed', 'latency', 'real-time applications', 'speculative decoding', 'Medusa', 'lookahead decoding', 'draft models'), though it leans technical and omits more colloquial phrasings like 'speed up my model' or 'make inference faster'.	4 / 5
Distinctiveness Conflict Risk	Occupies a clear niche built around three named, specific decoding techniques (speculative decoding, Medusa, lookahead decoding) with distinct triggers and minimal overlap risk against general inference or training skills.	5 / 5
	Total	18 / 20 Passed

Validation

93%

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 — 15 / 16 Passed

Validation for skill structure

Criteria	Description	Result
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning

	Total	15 / 16 Passed

Repository: synthetic-sciences/openscience
Path: backend/cli/skills/ml-inference/speculative-decoding/SKILL.md
Commit: 7dabbf0

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.