dcurlewis/.claude/skills/extract

Extract facts, decisions, and context changes from the current conversation that are worth persisting to memory.

Purpose

Conversations produce context that would otherwise be lost when the session ends. This command reviews the current conversation and writes a structured extraction file to be picked up by the normal memory pipeline.

This is a lightweight triage tool, not a comprehensive summary. The goal is to capture the 3-10 items from a conversation that would change or add to the memory files, not to document everything discussed.

When to Use

Run this at the end of a substantive conversation, particularly when:

• Decisions were made or positions clarified
• New facts emerged (people changes, project status shifts, vendor updates)
• Strategic context changed (priorities shifted, new constraints identified)
• Action items were agreed that aren't captured elsewhere
• Your thinking on a topic evolved in a way worth recording

Do NOT run this after:

• Routine drafting tasks (writing a Slack message, reviewing a doc)
• Simple factual lookups or questions
• Conversations that only produced an artefact already in the pipeline (e.g. a meeting summary)

Process

Environment Note

When running in Claude Desktop with project context, the memory files (memory-index.md, memory-context.md, memory-work.md, memory-people.md, about-me.md, reference.md) are already loaded as project documents. Use the in-context versions rather than reading them from disk. File writes (Step 4) still require filesystem tools regardless of environment.

Step 1: Review the Conversation

Scan the full conversation history in this session. Identify candidate items across three categories:

Facts & status changes (maps to memory-work.md or memory-context.md):

• Project status updates, new blockers, resolved issues
• Vendor or budget developments
• Technical decisions or architecture changes
• Deadline changes or delivery milestones

People & dynamics (maps to memory-people.md):

• Team health signals, retention concerns, morale shifts
• Relationship dynamics, alignment or friction patterns
• Individual development or performance observations
• Leadership interactions worth tracking

Strategic & organisational (maps to memory-context.md):

• Priority shifts or reframing
• New strategic context or constraints
• Organisational structure changes
• Planning or goal adjustments

Step 2: Apply Extraction Filters

For each candidate item, apply these filters:

1. Novelty: Is this genuinely new, or does it restate what's already in the memory files? Read memory-index.md to check current priorities and last update date. If an item merely confirms existing memory content without adding new detail, discard it.

2. Durability: Will this still matter in 1-2 weeks? Transient frustrations, speculative ideas that weren't committed to, or in-the-moment reactions generally don't qualify.

3. Actionability: Does this change what you or your teams should do, or how the memory files should represent the current state? If it doesn't change anything, it's not worth extracting.

4. Attribution: Is the source of this information clear? Extractions should note whether context came from the user directly, from a document discussed, from analysis during the conversation, etc. The memory pipeline needs to know the provenance.

5. Not already captured: If the conversation produced an artefact that will enter the pipeline independently (a meeting summary, a processed document, a Slack digest), don't duplicate those facts here. Only extract items that would otherwise be lost.

Step 3: Present Candidates to the User

Before writing anything, present the extracted items grouped by memory file mapping:

Proposed extractions from this session:

**memory-work.md**:
- [Item 1]
- [Item 2]

**memory-people.md**:
- [Item 3]

**memory-context.md**:
- [Item 4]

**Discarded** (didn't pass filters):
- [Item 5] — reason: already in memory
- [Item 6] — reason: speculative, no commitment made

Ask the user to confirm, remove, or add items. This step is mandatory; do not skip it.

Step 4: Write Extraction File

After the user confirms, write the extraction to:

Raw-Materials/Session-Extracts/YYYYMMDD-session-extract.md

If multiple extractions happen on the same day, append a sequence suffix: YYYYMMDD-session-extract-2.md.

Create the Raw-Materials/Session-Extracts/ directory if it does not exist.

Use this format:

---
kind: session-extract
authorship: ai-assisted
date: YYYY-MM-DD
session_context: [Brief 1-line description of what the conversation was about]
---

# Session Extract — YYYY-MM-DD

## Context

[1-2 sentences describing the conversation topic and why these items were extracted.]

## Extractions

### memory-work.md

- **[Bold topic]**: [Concise factual statement of what changed or was decided]
- **[Bold topic]**: [Another item]

### memory-people.md

- **[Bold topic]**: [Concise factual statement]

### memory-context.md

- **[Bold topic]**: [Concise factual statement]

Omit any section heading that has no items beneath it.

Writing Standards

• Each extraction item should be 1-2 sentences maximum
• State facts, not interpretations. "User expressed concern about X" is fine; "User is worried about X" is interpretation
• Use the same factual, non-dramatic tone as the memory files
• Include specific names, dates, and numbers where available
• Do not editoralise or add analysis; the memory pipeline will contextualise during consolidation

What NOT to Extract

• Drafts of messages or documents (the artefact itself is the output, not memory)
• Claude's recommendations or analysis (unless the user explicitly endorsed them as a decision)
• Routine task completion ("we drafted a Slack message to a colleague")
• Emotional reactions or venting (unless they signal a durable dynamic worth tracking)
• Speculative scenarios that weren't committed to
• Information that came from reading existing memory files (that's already captured)

Pipeline Integration

Session extracts land in Raw-Materials/Session-Extracts/ and are picked up by /memory-scan during the next memory update cycle. They are treated as a source file like any other raw material.

Output

1. Write the extraction file to Raw-Materials/Session-Extracts/YYYYMMDD-session-extract.md
2. Report to the user:
   • Number of items extracted
   • File location
   • Reminder: items will be picked up by the next /memory-scan run