Make Distilled

Make Distilled Content

When to use

Use this skill when:

• The user has a raw file in raw/ and wants to extract structured insights
• You need to convert unprocessed source material into organized knowledge
• You're preparing content for integration into notes or essays

Keywords: distill, extract, summarize, patterns, principles, organize

Inputs

Required:

• raw_paths (list[string]): Relative or absolute paths to one or more raw files in raw/

Optional:

• focus (string): Specific aspect to emphasize (e.g., "patterns", "principles", "practical advice")
• distilled_filename (string): Output filename (useful when distilling multiple raw files into one distilled file)

Outputs

This skill produces:

1. A new file in distilled/ with the same timestamp-slug naming pattern as the raw file
2. An update to the raw file front matter to add a backlink to one (or more) distilled outputs
3. Metadata returned to the agent:
   • distilled_path: full path to the created distilled file
   • title: title of the distilled content
   • distilled_at: ISO timestamp (UTC)
   • status: initial status (typically "draft")

Procedure

1. Read the raw files

• Use the read_file tool to load each raw file content
• Parse front matter from each raw file to extract:
  • title
  • source_url
  • captured_at
  • capture_type
  • author (if available)
  • published_at (if available)
  • distilled_refs (if present; used for appending a new backlink)
  • any other relevant metadata

2. Generate distilled content structure

Extract and organize the following elements from the raw content:

Summary (5-10 lines)

• Concise overview of the main topic and purpose
• What problem/question does this address?
• What approach or solution is presented?

Key points (bullet list)

• 5-10 actionable or memorable insights
• Focus on practical takeaways
• Avoid restating the summary

Concepts / principles

• Abstract ideas or mental models
• Transferable patterns
• Underlying theory or philosophy
• Trade-offs and decision factors

Patterns (if relevant)

• Concrete, reusable approaches
• Workflow patterns
• Code/architectural patterns
• Team/organizational patterns

Entities (if relevant)

• Tools, frameworks, libraries mentioned
• People or organizations referenced
• Related projects or initiatives
• Terminology and definitions

Quotes (select 3-8 notable ones)

• Verbatim excerpts that are particularly insightful
• Include just enough context to be standalone
• Format as blockquotes with inline attribution if author is known

Open questions / follow-ups

• Gaps or areas for deeper research
• Connections to explore
• Uncertainties or ambiguities in the source

Next steps (agent-suggested)

• Suggested integration targets in notes/
• Related topics to capture or research
• Concepts worth expanding into standalone notes

3. Build front matter

Create YAML front matter for the distilled file:

---
title: "<original title> (or improved title if clearer)"
source_url: "<primary source URL (optional)>"
captured_at: "<captured_at from a representative raw file>"
distilled_at: "<current ISO timestamp UTC>"
raw_refs:
  - "[[raw/<raw file 1>]]"
  - "[[raw/<raw file 2>]]"
capture_type: "<capture_type from a representative raw file>"
status: draft
agent: github-copilot
model: claude-sonnet-4.5
confidence_notes: "<any uncertainty flags, hallucination risks, or missing context>"
tags: ["tag1", "tag2", "tag3"]  # 3-7 relevant tags
---

Critical rules:

• Use raw_refs as an Obsidian-style internal link list (e.g., [[raw/20260102-095107Z--patterns-for-ai-assisted-software-development]]).
• Always use raw_refs (plural) as a list, even when there is only one raw source.
• Also add (or append to) a backlink in each raw file front matter:
  • distilled_refs should be a YAML list of Obsidian-style internal links (e.g., [[distilled/20260102-095107Z--patterns-for-ai-assisted-software-development]])
• In YAML front matter, quote Obsidian links (e.g., - "[[raw/...]]") so the file stays valid YAML.
• Be explicit in confidence_notes if:
  • Author/date was inferred rather than explicit
  • Content was truncated or unclear
  • You're uncertain about any extraction
• The status should be draft unless the user specifies otherwise

4. Apply the distilled content template

Combine the extracted elements into this structure:

---
<front matter from step 3>
---

## Summary

<5-10 line summary>

## Key points

- <point 1>
- <point 2>
- ...

## Concepts / principles

<extracted concepts, one per paragraph or sub-section>

## Patterns

<if relevant: concrete patterns found>

## Entities

<if relevant: tools, people, projects, terms>

## Quotes

> <quote 1>
> 
> — <attribution if known>

> <quote 2>

...

## Open questions / follow-ups

- <question or gap 1>
- <question or gap 2>

## Next steps

- <suggested action 1>
- <suggested action 2>

## Links

- Source: [<source_url>](<source_url>)
- (Optional) Additional external links extracted from the source (avoid duplicating `raw_refs` here; they live in front matter)
- <Other links extracted from content if relevant>

5. Generate filename and write the file

• If distilled_filename is provided, use it.
• Otherwise:
  • If raw_paths contains exactly one file, use the same timestamp and slug as that raw file
  • If raw_paths contains multiple files, generate a new YYYYMMDD-HHMMSSZ--<slug>.md filename based on the distilled title
• Full path: distilled/<filename>
• Content: front matter + blank line + distilled structure
• Use create_file tool with the full content

6. Update raw front matter (backlink)

• For each file in raw_paths:
  • Keep the raw body unchanged; only update front matter.
  • Add or update distilled_refs as a YAML list:
    • If distilled_refs is missing, add it.
    • If distilled_refs exists, append the new distilled path (avoid duplicates).
  • Use Obsidian-style internal links, e.g.:

distilled_refs:
  - "[[distilled/20260102-095107Z--patterns-for-ai-assisted-software-development]]"

7. Confirm to user

Provide a brief confirmation:

• Link to the created distilled file using an Obsidian-style internal link
• One-sentence summary of what was extracted
• Note any significant confidence_notes or gaps

Quality guidelines

Preserve intent and nuance

• Do not "improve" the source's tone or claims
• If something is speculative in the source, preserve that uncertainty
• Avoid marketing language or hype

Be explicit about uncertainty

• Use confidence_notes front matter for extraction risks
• In the body, flag inferred vs. explicit information
• If author/date is missing, say so

Favor principles over tools

• Extract transferable ideas, not tool-specific instructions
• Highlight trade-offs and decision factors
• Patterns should be adaptable across contexts

Short, precise, reusable

• Clarity over verbosity
• Make each section independently useful
• Link concepts instead of duplicating them

Avoid exposing private information

• Do not include private repository names, local paths, or personal data
• Sanitize examples if they reference non-public systems

Examples

Example 1: Blog post about AI agent patterns

Input:

raw_paths:
  - raw/20260102-095107Z--patterns-for-ai-assisted-software-development.md

Process:

1. Read the raw file and extract metadata
2. Identify main patterns: interview-driven specs, progressive disclosure, etc.
3. Extract key principles: preserve intent, think in systems, avoid hype
4. Select 5-6 notable quotes
5. Note any follow-up questions (e.g., "How do these patterns scale to larger teams?")

Output:

• Creates distilled/20260102-095107Z--patterns-for-ai-assisted-software-development.md
• Front matter includes raw_refs: ["[[raw/20260102-095107Z--patterns-for-ai-assisted-software-development]]"]
• Raw front matter updated with distilled_refs: ["[[distilled/20260102-095107Z--patterns-for-ai-assisted-software-development]]"]
• Structured summary + key points + concepts + quotes + next steps

Example 2: YouTube transcript

Input:

raw_paths:
  - raw/20260105-140000Z--building-production-agents.md

Process:

1. Read the transcript
2. Extract main themes (often less linear than written content)
3. Identify practical advice vs. theoretical discussion
4. Note any tools or frameworks mentioned
5. Flag areas where transcript was unclear or incomplete

Output:

• Creates distilled/20260105-140000Z--building-production-agents.md
• confidence_notes might mention: "Transcript had several unclear segments; some technical terms may be misspelled"
• Entities section lists tools and frameworks discussed
• Raw front matter updated with distilled_refs (append if already present)

Failure modes and edge cases

Truncated or incomplete raw content

• Note in confidence_notes
• Extract what's available
• Suggest follow-up: "Consider re-capturing with full content"

Highly technical content with domain-specific jargon

• Preserve technical terms as-is
• Add brief context in Entities section if helpful
• Do not oversimplify at the cost of accuracy

Opinion pieces vs. technical guides

• For opinion: focus on arguments, not just conclusions
• For guides: emphasize patterns and principles over step-by-step

Multiple topics in one source

• Consider suggesting multiple distilled files focused on different aspects
• Or use clear sub-sections in a single distilled file

Tools typically used

• read_file — to load the raw content
• create_file — to write the distilled output
• update_file (or equivalent) — to update raw front matter with distilled_refs
• run_in_terminal — to generate UTC timestamp if needed

References

• Distillation pipeline: docs/distillation/distillation-pipeline.md
• Agent skills standard: notes/agent-skills.md
• Repository guidance: AGENTS.md