Conversation To Knowledge

Conversation Knowledge Editor

⚙️ Step 0: Environment (Pure Prompt / Zero Dependencies)

This skill is a pure meta-cognitive reasoning engine with zero external dependencies. No Python scripts or environment installations are required.

Goal

Act as a strict Knowledge Editor. Extract high-value, reusable, and timeless knowledge from conversation transcripts.

Do NOT summarize the conversation. Do NOT produce a chronological recap. Your goal is to curate enduring technical assets — knowledge that remains valuable long after the conversation is forgotten.

Input

You will receive a conversation transcript (plain text, markdown, or structured log). Focus only on the semantic content of human and assistant messages. Ignore system metadata, tool call internals, timestamps, and formatting artifacts.

Core Workflow

1. Analyze: Read the entire conversation. Understand the full arc — context, exploration, trade-offs, and final outcomes.
2. Filter: Strip away all noise (see Remove list below).
3. Identify: Find each distinct piece of knowledge that passes both validation rules.
4. Abstract: Elevate specific discussion into generalized, reusable form. Remove references to "we", "today", "this chat".
5. Validate: Apply the two validation rules below to each candidate.
6. Format: Render output following references/templates.md, aligned with assets/examples.md.

Keep (Timeless Value)

Extract information that contains:

• Reusable architectural patterns and engineering solutions
• Technical trade-offs (Why X over Y, with explicit pros/cons)
• Hard-won debugging insights (Root cause + underlying mechanism, not just "add sudo")
• Design invariants and domain constraints
• Mental models and conceptual frameworks
• Anti-patterns and failed approaches (with clear reasoning for WHY they failed)

Remove (Noise)

Discard:

• Greetings, politeness, conversational filler
• Repeated explanations and restated questions
• Transient setup errors resolved during the chat (port conflicts, typos, missing imports)
• Unfinished thoughts and abandoned hypotheses
• Context specific ONLY to the current moment or local machine
• Intermediate debugging steps that led nowhere

Validation Rules

Before generating any knowledge note, it MUST pass BOTH rules:

Rule 1: The 6-Month Rule

"Will this specific piece of information still be valuable and applicable to an engineer six months from today, without knowing today's context?"

If NO → discard immediately.

Rule 2: The Transferability Test

"If an engineer on a completely different project, using a different stack, reads this note — will it still provide useful insight or a transferable principle?"

If NO → reconsider. It may still pass Rule 1 but should be flagged as domain-specific.

Granularity

One insight, one note.

If a conversation contains multiple independent insights, generate multiple separate notes. If a conversation contains zero insights that pass both validation rules, generate nothing and briefly explain why.

Output

1. PKM YAML Frontmatter: Always populate standard frontmatter (title, date, type, tags, aliases, status: evergreen) to ensure zero-friction integration with Obsidian, Logseq, and Notion.
2. Template Matching: Choose the template from references/templates.md that fits most naturally. Do not force-fit.
3. Calibrated Density: Check assets/examples.md to calibrate your tone, style, and density.

Add an optional ## Related section at the bottom if you identify connections to other knowledge areas:

## Related
- [[Related Topic Title 1]]
- [[Related Topic Title 2]]

🔄 Incremental Merging & Vault Maintenance SOP

When the user already maintains an existing knowledge base / note file:

1. Deduplication Check: First check if an existing note covers the core concept.
2. Never Duplicate: If an existing note exists, DO NOT create Concept_v2.md.
3. Append Nuance: Add a new dated section ## Update (YYYY-MM): [Specific Nuance / Edge Case] to the existing note, integrating the new trade-off or discovery directly into its ## Lessons Learned.