Polish Transcriptions Skill

Transform raw, machine-generated transcriptions into polished, cognitively-ordered Obsidian notes that are both readable and complete.

Objective

Convert poorly transcribed audio/video content (workshops, lectures, meetings, interviews) into well-structured, publication-ready documents while preserving 100% of the original information.

Core Principles

1. Zero Information Loss

[!danger] Critical Requirement Never omit, summarize, or compress information from the original. Every detail, example, tangent, question, and answer must be preserved. The output should contain MORE structure, not LESS content.

2. Cognitive Reorganization

Transform stream-of-consciousness speech into logical document sections:

Speech Pattern Transforms To

Topic jumping Grouped sections with headers

Repetition Single consolidated statement

Filler words/false starts Clean prose

Tangents Callouts or integrated context

Q&A interruptions Blockquote dialogues

3. Semantic Structure Over Chronological Order

Reorganize content by meaning, not by when things were said. A 2-hour rambling lecture about three topics becomes three clean sections, even if the speaker jumped between them.

Transformation Process

Phase 1: Analysis

Before writing anything:

• Read the entire transcript — Understand all topics covered

• Identify main themes — What are the 3-7 core topics?

• Categorize content types:

• Core instruction/information

• Examples and anecdotes

• Q&A interactions

• Meta-commentary (jokes, digressions)

• Action items or recommendations

• Map relationships — Which topics depend on others?

Phase 2: Structure Design

Create a logical outline:

[Main Topic 1]

[Subtopic 1.1]

[Subtopic 1.2]

[Main Topic 2]

...

Use horizontal rules (--- ) to separate major topic shifts.

Phase 3: Content Transformation

Apply these transformations systematically:

Headers and Hierarchy

Main Section <!-- H2 for major topics -->

Subsection <!-- H3 for subtopics -->

Point or Example <!-- H4 for specific items when needed -->

Dialogues and Q&A

Preserve speaker identities with blockquotes:

Participante: ¿Cómo funciona X? Instructor: X funciona de esta manera...

For multi-turn exchanges:

Estudiante: Primera pregunta Profesora: Respuesta inicial Estudiante: Pregunta de seguimiento Profesora: Respuesta expandida

Callouts for Special Content

Content Type Callout to Use

Key concept/principle

[!important]

Practical advice

[!tip] Recomendación

Warning/caution

[!warning]

Interesting aside

[!note]

Real-world example

[!example]

Quoted wisdom

[!quote]

Action items

[!todo]

Summary

[!abstract] or > [!tldr]

Success/conclusion

[!success]

Tables for Structured Data

Convert comparison discussions into tables:

Lists for Enumerated Content

When the speaker lists things (even implicitly):

• Item one
• Item two
  • Sub-item
• Item three

Mermaid Diagrams for Processes

When a process or flow is described:

graph LR
    A[Paso 1] --> B[Paso 2]
    B --> C[Paso 3]
    C --> D[Resultado]

Code Blocks for Technical Content

# Example code from the presentation
def example():
    return "formatted code"

Formatting Standards

Frontmatter

Always include appropriate YAML frontmatter:

date: YYYY-MM-DD professor: "[[Speaker Name]]"

or

speaker: "[[Speaker Name]]"

optional

tags:

• workshop
• topic

Text Formatting

Purpose Syntax Example

Key terms first mention bold

machine learning

Technical terms code

SQL

Emphasis italic

very important

Highlighting ==text==

==critical deadline==

Links

Create wikilinks for concepts that deserve their own notes:

Esto se relaciona con [[machine learning]] y [[data science]].

Anti-Patterns (What NOT To Do)

❌ Summarizing

<!-- BAD: Lost information -->

El instructor habló sobre varios temas de datos.

<!-- GOOD: Preserves detail -->

El instructor cubrió tres áreas principales:

1. Integración de datos — consolidar información de múltiples fuentes
2. Limpieza y transformación — ordenar, depurar y preparar los datos
3. Análisis exploratorio — comprender patrones y comportamientos

❌ Removing "Unimportant" Content

<!-- BAD: Removes color and context -->

(omitted anecdote about COVID impact)

<!-- GOOD: Preserves as callout -->

[!example] Caso Real: El Impacto del COVID-19 En un banco donde trabajé, teníamos modelos de predicción de mora...

❌ Flattening Dialogue

<!-- BAD: Loses attribution -->

Se discutió que SQL es el lenguaje principal.

<!-- GOOD: Preserves interaction -->

Estudiante: ¿Qué es SQL? Profesora: SQL es el lenguaje de programación de bases de datos.

❌ Over-Structuring

<!-- BAD: Too many headers for simple content -->

Definición de Dato

Tipo 1

Subtipo A

<!-- GOOD: Appropriate nesting -->

Tipos de Datos

• Tipo 1: Descripción
  • Subtipo A

Quality Checklist

Before delivering the polished document:

• Information complete — All original content is present

• Logical structure — Grouped by topic, not chronology

• Frontmatter present — Date, speaker/professor, optional tags

• Headers used correctly — H2 for sections, H3 for subsections

• Dialogues preserved — Q&A in blockquote format with speaker names

• Callouts appropriate — Important points in [!tip] , [!important] , etc.

• Tables where helpful — Comparisons and structured data formatted

• Mermaid diagrams — Processes visualized when described

• Bold for key terms — First mention of important concepts

• Wikilinks created — Concepts linked with [[concept]]

• Horizontal rules — Major topic separations marked with ---

• Clean prose — No filler words, false starts, or transcription artifacts

• No orphan headers — Every header has content below it

Example Transformation

Before (Raw Transcription)

bueno entonces ehh vamos a ver lo de las bases de datos entonces una base de datos es pues como un lugar donde guardas cosas no? ah esperen me olvidé de decirles mi nombre soy Carmen ehh entonces como les decía hay diferentes tipos de bases de datos algunas son relacionales otras no relacionales las relacionales usan SQL que es un lenguaje de programación bueno no exactamente programación pero sirve para consultar datos entonces SQL significa structured query language y sirve para hacer consultas a la base de datos...

After (Polished Document)

date: 2025-08-08 professor: "[[Carmen Marín]]"

Introducción a las Bases de Datos

Una base de datos es un almacén centralizado donde se guardan y organizan datos para su posterior acceso y manipulación.

Tipos de Bases de Datos

SQL (Structured Query Language)

SQL es el lenguaje estándar para interactuar con bases de datos relacionales. Permite realizar consultas, inserciones, actualizaciones y eliminaciones de datos.

[!note] Aclaración Aunque SQL contiene elementos de programación, técnicamente es un lenguaje de consulta, no un lenguaje de programación de propósito general.

Workflow Integration

Suggested Process

• Read the obsidian-markdown skill first for syntax reference

• Analyze the complete raw transcript

• Outline the logical structure

• Transform section by section

• Review against the quality checklist

• Verify no information was lost by comparing key facts

Output Location

Polish transcriptions should be saved to the appropriate location in the user's vault, typically:

• 03 resources/ for workshops and external content

• 01 projects/.../classes/ for academic lectures

• Same directory as source with a new filename

Success Criteria

A successfully polished transcription:

• Reads like a well-written article — Not like speech

• Contains all original information — Nothing omitted

• Uses Obsidian features effectively — Callouts, tables, diagrams

• Has clear cognitive structure — Easy to navigate and reference

• Preserves speaker personality — Quotes and dialogues maintain voice

• Is immediately usable — No further editing needed by user

References

• Obsidian Callouts Documentation

• Obsidian Properties and Frontmatter

• Mermaid Diagram Syntax

• CommonMark Specification