Knowledge Base Skill

Research and document technical knowledge with structured semantic metadata in Obsidian.

Note types:

• •Introduction — Concept primer for someone encountering the topic for the first time
• •Deep Dive — Advanced internals, mechanisms, edge cases, failure modes
• •Comparison — Side-by-side analysis of 2+ concepts to support a decision
• •Tutorial — Step-by-step practical guide for accomplishing a task
• •Survey — Landscape overview mapping an entire field or category

Workflow

1. Determine Note Type

Use this decision tree:

If the user doesn't specify, default to Introduction. Ask only if genuinely ambiguous.

2. Research

Use WebSearch to gather comprehensive information:

WebSearch: "[concept] definition explanation"
WebSearch: "[concept] examples use cases"
WebSearch: "[concept] advantages disadvantages limitations"
WebSearch: "[concept] vs [related concept]" (for counterparts)
WebSearch: "[concept] history origin year introduced paper"
WebSearch: "[concept] governing principles design criteria" (what laws/principles govern this concept?)

IMPORTANT: Always research:

• •When the concept was first introduced (year, original paper/author)
• •Sophistication level (basic building block vs cutting-edge)
• •Governing principles or laws that constrain the concept (e.g., Goodhart's Law for benchmarks)

3. Extract Semantic Relationships

From research, identify:

• •Synonyms: Alternative names (e.g., "cost function" = "loss function")
• •Antonyms: True opposites (e.g., "overfitting" ↔ "underfitting")
• •Hypernyms: Broader categories (IS-A relationship, e.g., "cross-entropy" IS-A "loss function")
• •Hyponyms: More specific variants (e.g., "binary cross-entropy" is a type of "cross-entropy")
• •Counterparts: Related concepts often compared (e.g., "precision" & "recall")

4. Create the Note

Read the appropriate template from references/ and create the note:

Frontmatter Templates

Concept (Introduction + Deep Dive)

---
fileClass: Concept
note_type: Introduction  # or Deep Dive
domain: [ML/AI | NLP | Deep Learning | Statistics | Mathematics | Computer Science | General]
year_introduced: YYYY
sophistication: [Foundational | Intermediate | Advanced | State-of-the-Art]
antonyms:
  - opposite_concept
synonyms:
  - alternative_name
hypernyms:
  - broader_category
hyponyms:
  - specific_variant
counterparts:
  - comparison_concept
related_concepts:
  - "[[Related Note]]"
date_created: YYYY-MM-DD
date_modified: YYYY-MM-DD
tags:
  - concept
  - [domain-tag]
---

Comparison

---
fileClass: Comparison
concepts_compared:
  - concept_a
  - concept_b
domain: [ML/AI | NLP | Deep Learning | Statistics | Mathematics | Computer Science | General]
verdict: "Brief recommendation or conclusion"
related_concepts:
  - "[[Related Note]]"
date_created: YYYY-MM-DD
date_modified: YYYY-MM-DD
tags:
  - comparison
  - [domain-tag]
---

Tutorial

---
fileClass: Tutorial
domain: [ML/AI | NLP | Deep Learning | Statistics | Mathematics | Computer Science | General]
difficulty: [Beginner | Intermediate | Advanced]
prerequisites:
  - "[[Prerequisite Concept]]"
estimated_time: "30 minutes"
related_concepts:
  - "[[Related Note]]"
date_created: YYYY-MM-DD
date_modified: YYYY-MM-DD
tags:
  - tutorial
  - [domain-tag]
---

Survey

---
fileClass: Survey
domain: [ML/AI | NLP | Deep Learning | Statistics | Mathematics | Computer Science | General]
scope: "Brief description of what this survey covers"
num_concepts: 0
related_concepts:
  - "[[Related Note]]"
date_created: YYYY-MM-DD
date_modified: YYYY-MM-DD
tags:
  - survey
  - [domain-tag]
---

Content Quality Guidelines

Definition Quality

• •Identify the fundamental tension the concept addresses (e.g., a benchmark's tension is between "measuring real capability" and "being practical to compute")
• •Surface proxy relationships — what does this concept approximate, and what gets lost? (e.g., "a benchmark score is a proxy for real-world capability")
• •Keep to 2-3 sentences. Move elaboration to other sections.

Governing Principles

After Key Characteristics, include a Governing Principles section when a known law or principle constrains the concept:

> [!abstract] Goodhart's Law
> **"When a measure becomes a target, it ceases to be a good measure."**
> — Charles Goodhart (1975)
>
> [Explanation of how this principle specifically governs the concept being documented]

Organize the Known Limitations section under the governing principles — limitations should flow from the principles, not be a flat list.

Design Criteria (Conditional)

Include a "Design Criteria" section only for tools, methods, and frameworks — things that can be designed well or poorly. Ask: "What makes a good X?" Skip for pure theoretical concepts.

Content Freshness

• •Prefer patterns over perishable specifics (say "transformer-based models" not "GPT-4")
• •Never hardcode model version numbers or leaderboard rankings
• •Use relative timeframes ("as of writing" or "since 2020") instead of absolute claims about "the latest"
• •Facts that will age should be in Examples, not in Definition

Comparison Tables

• •Only compare methodologically parallel things in the same table
• •Bad: comparing a metric with a benchmark suite in the same row
• •Good: comparing two metrics, or two benchmark suites, against the same aspects
• •Every column header should be the same "kind of thing"

Section Order (Concept Notes)

The preferred section order for Concept notes (Introduction + Deep Dive) is:

1. •Definition
2. •Concept Family Tree
3. •Comparison & Trade-offs
4. •Historical Context & Sophistication
5. •Key Characteristics
6. •Governing Principles (conditional)
7. •Design Criteria (conditional)
8. •Known Limitations
9. •Examples
10. •Internals / Mechanism (Deep Dive only)
11. •Mathematical Formulation (conditional)
12. •Use Cases
13. •Related Concepts
14. •References

Concept Family Tree

• •Place the concept at the correct taxonomic level — don't make it appear as a leaf if it's actually a branch
• •Never make the concept appear as its own child or parent (no self-referential loops)
• •Include at least 2 levels above and 1 level below the concept

Related Concepts

• •Aim for 5+ wikilinks in the Related Concepts section
• •Link every concept mentioned in the body that has or could have its own note
• •Use wikilink format: [[Concept Name]]

Conditional Sections

• •Mathematical Formulation: Only include for concepts with core formulas. Skip for high-level concepts, tools, or frameworks.
• •Design Criteria: Only for tools/methods/frameworks (see above)
• •Examples: Use the most natural format — narrative, calculation, before/after, code snippet. Don't force the rigid Setup/Calculation/Result structure.

Reference Templates

Read the appropriate template when creating a note:

• •Introduction / Deep Dive: Read references/concept-template.md
• •Comparison: Read references/comparison-template.md
• •Tutorial: Read references/tutorial-template.md
• •Survey: Read references/survey-template.md

Each template is self-contained with full body structure, section guidance, and FileClass reference.

Sophistication Levels Guide

Notes

• •Always use WebSearch for current, accurate information
• •ALWAYS include the year the concept was introduced — research the original paper/author
• •ALWAYS explain the sophistication level — Is it a basic building block or cutting-edge?
• •Include numerical examples whenever possible for ML/Statistics concepts
• •Use LaTeX for mathematical formulations
• •Create wikilinks to existing notes in the vault when known
• •Keep the definition section concise; use other sections for details