Ml Primitive Decoder

Decomposes any ML construct (attention, layer norm, softmax, convolution, dropout, contrastive loss, diffusion step, gradient descent update, cross-entropy) into the small set of linear-algebra primitives it's built from, then explains why those primitives produce the observed behavior. Includes an ablation thought experiment that asks "what would break if you removed this piece?". Use when the user asks "why does X work?", "explain attention/conv/norm intuitively", "what's actually happening in this layer", or when reading an architecture paper and a particular block doesn't make sense yet.

200 stars 25 forksAI, LLMs & Agents Plugin marketplace

Install

Claude Code plugin
/plugin marketplace add lyndonkl/claude

The repo is a Claude Code plugin marketplace. Run /plugin afterwards to pick the plugin that includes this skill.

skills CLI (Claude Code, Cursor, Codex and more)
npx skills add lyndonkl/claude --skill ml-primitive-decoder
Manual install
git clone --depth 1 https://github.com/lyndonkl/claude.git /tmp/claude
mkdir -p ~/.claude/skills
cp -r /tmp/claude/skills/ml-primitive-decoder ~/.claude/skills/ml-primitive-decoder

Personal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.

From the SKILL.md

# ML Primitive Decoder

## Table of Contents
1. [Workflow](#workflow)
2. [The Primitive Catalog](#the-primitive-catalog)
3. [Decoding Patterns](#decoding-patterns)
4. [Ablation Thought Experiment](#ablation-thought-experiment)
5. [Common Patterns](#common-patterns)
6. [Guardrails](#guardrails)
7. [Quick Reference](#quick-reference)

Most ML constructs that look opaque are actually a short stack of well-understood linear algebra primitives composed in a specific way. Attention is dot product + softmax + weighted sum. Layer norm is recenter + rescale + learnable affine. Diffusion is noise injection + denoising regression repeated. This skill names the primitives, then shows why their composition produces the behavior the construct is famous for.

The signature deliverable is: **construct = primitive₁ + primitive₂ + primitive₃, and *that's why* it does X**.

**Quick example (Attention):**

> **Construct:** Self-attention layer, output = softmax(QKᵀ/√d) · V.
>
> **Primitives:**
> 1. **Linear projections** (Q, K, V each = a different learnable linear layer applied to the same input).
> 2. **Dot product** (QKᵀ measures pairwise similarity between query and key vectors).
> 3. **Scaled softmax** (divides scores by √d to keep variance bounded, then row-normalizes into mixing weights).
> 4. **Weighted sum** (multiplying by V is a per-row weighted average over value vectors).
>
Read the full skill on GitHub

Details

Skill name
ml-primitive-decoder
Path
skills/ml-primitive-decoder/SKILL.md
Repo updated
1 September 2026

FAQ

What does the Ml Primitive Decoder skill do?

Decomposes any ML construct (attention, layer norm, softmax, convolution, dropout, contrastive loss, diffusion step, gradient descent update, cross-entropy) into the small set of linear-algebra primitives it's built from, then explains why those primitives produce the observed behavior. Includes an ablation thought experiment that asks "what would break if you removed this piece?". Use when the user asks "why does X work?", "explain attention/conv/norm intuitively", "what's actually happening in this layer", or when reading an architecture paper and a particular block doesn't make sense yet.

How do I install the Ml Primitive Decoder skill?

Claude Code plugin: /plugin marketplace add lyndonkl/claude. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.

Is Ml Primitive Decoder popular?

It lives in lyndonkl/claude, which has 200 GitHub stars and 25 forks.

Related skills

Abstraction Concrete Examples

200

lyndonkl/claude

Builds structured abstraction ladders that translate high-level principles into concrete, actionable examples across 3-5 levels. Bridges communication gaps, reveals hidden assumptions, and tests whether abstract ideas work in practice. Use when explaining concepts at different expertise levels, moving between abstract principles and concrete implementation, identifying edge cases by testing ideas against scenarios, designing layered documentation, decomposing complex problems into actionable steps, or bridging strategy-execution gaps.

Testing & QA25

Academic Letter Architect

200

lyndonkl/claude

Guides the creation of evidence-based academic recommendation letters, reference letters, and award nominations that combine concrete examples, meaningful comparisons, and genuine enthusiasm. Use when writing recommendation letters for students, postdocs, or colleagues, or when user mentions recommendation letter, reference, nomination, letter of support, endorsement, or needs help with strong advocacy and comparative statements.

Research & Science25

Adr Architecture

200

lyndonkl/claude

Documents significant architectural and technical decisions with full context, alternatives considered, trade-offs analyzed, and consequences understood. Creates a decision trail that helps teams understand why decisions were made. Use when choosing between technology options, making infrastructure decisions, establishing standards, migrating systems, or when user mentions ADR, architecture decision, technical decision record, or decision documentation.

Software Development25

Mlb Decision Logger

200

lyndonkl/claude

Appends structured decision entries to the yahoo-mlb decision log (tracker/decisions-log.md) on behalf of any agent in the MLB team. Validates entries against the authoritative schema, serializes concurrent writes from parallel agents, and runs the Monday calibration pass to fill in outcomes and update the variant scoreboard. Use when any MLB agent needs to record a decision, when the coach requests "log decision", "append to decision log", "record variant outcome", or runs the "calibration pass".

AI, LLMs & Agents25

Meta Prompt Engineering

200

lyndonkl/claude

Transforms vague or unreliable prompts into structured, constraint-aware prompts with explicit roles, task decomposition, output formats, and quality checks. Use when prompts produce inconsistent outputs, need explicit structure and constraints, require safety guardrails, involve multi-step reasoning that needs decomposition, need domain expertise encoding, or when user mentions improving prompts, prompt templates, structured prompts, prompt optimization, reliable AI outputs, or prompt patterns.

AI, LLMs & Agents25

Mlb Faab Sizer

200

lyndonkl/claude

Computes FAAB (Free Agent Acquisition Budget) recommended and maximum bids for Yahoo fantasy baseball waiver targets. Implements the baseball-specific layering of the faab-bid-framework (positional_need_fit, role_certainty, urgency, season_pace, league-inflation calibration) and DELEGATES the game-theoretic primitives -- first-price shading and winner's-curse haircut -- to the sibling skills `auction-first-price-shading` and `auction-winners-curse-haircut`. Produces a recommended bid, a hard ceiling, a rationale with the full delegation chain, and guardrail flags. Use when the user asks "how m

AI, LLMs & Agents25

Browse the most popular Claude skills, all ai, llms & agents skills, or write your own with the skills generator.