clotho-extraction

In-Session Extraction from Transcripts

Clotho's extraction pipeline runs within Claude sessions — you ARE the extraction engine. When a user provides a transcript or meeting notes, analyze the content and create structured entities.

Speech Act Ontology

Look for these patterns in transcripts:

Speech Act	Signal Patterns	Create As
Commit	"I'll do X", "I can take that", "Let me handle"	Task (speaker owns)
Decide	"We're going with X", "Decision is Y", "We've decided"	Decision
Risk	"The concern is...", "Risk here is...", "I'm worried about"	Risk
Block	"We're stuck on...", "Blocked by...", "Can't proceed until"	Blocker
Question	"We need to figure out...", "Open question:", "How do we..."	Question
Insight	"What we learned...", "Key takeaway...", "Interesting finding"	Insight
Delegate	"Can you take this?", "Assigning to X"	Task (target owns)
Request	"I need X from you", "Can you get me..."	Task (inbound)
Follow-up	"Let's schedule a meeting about...", "Set up time for...", "Let's follow up on...", "Circle back on..."	Task
Next steps	"Next steps:", "Action items:", "By next week we need..."	Task(s)

Before Starting

Check the extraction queue:

clotho_list_unprocessed()

If there are multiple unprocessed transcripts, ask the user:

"There are N unprocessed transcripts. Want to process them one at a time, or all at once?"

Default to one at a time for 5+ transcripts — batch extraction tends to overwhelm with questions.

Skip patterns

Check for .clotho/extraction-config.toml in the workspace. If it exists, read it for skip patterns:

[extraction]
# Transcripts matching these title patterns are auto-skipped (glob syntax)
skip_titles = ["*1:1*", "*sport*", "*standup - DS Infra*"]
# Mark skipped transcripts as processed so they don't reappear in the queue
mark_skipped = true

When processing, if a transcript title matches any skip pattern, skip it and (if mark_skipped = true) auto-mark as processed. Report skipped items in the summary.

If the user tells you to skip certain meeting types during the session, offer to add them to the config:

"Want me to add 'sport' to your extraction skip patterns so it's automatic next time?"

Extraction Workflow

1. Read the transcript — Capture it first if not already in the workspace:

   clotho_capture(file_path: "transcript.md", entity_type: "transcript", title: "Sprint Planning")
   

2. Identify speech acts — Read through the content and identify each speech act.

3. Create entities — For each identified item:

   clotho_create_entity(entity_type: "decision", title: "Go with microservice approach")
   clotho_create_entity(entity_type: "risk", title: "Database migration complexity")
   clotho_create_entity(entity_type: "task", title: "Write migration RFC")
   

4. Create relations — Link extracted entities back to the transcript. Use batch tool for efficiency:

   clotho_batch_create_relations(relations: [
     {"source_id": "<decision_id>", "relation_type": "extracted_from", "target_id": "<transcript_id>"},
     {"source_id": "<task_id>", "relation_type": "extracted_from", "target_id": "<transcript_id>"},
     {"source_id": "<task_id>", "relation_type": "belongs_to", "target_id": "<program_id>"}
   ])
   

5. Link people — Create Person entities and MENTIONS relations:

   clotho_create_entity(entity_type: "person", title: "Alice")
   clotho_batch_create_relations(relations: [
     {"source_id": "<transcript_id>", "relation_type": "mentions", "target_id": "<person_id>"}
   ])
   

6. Mark processed — After extraction is complete:

   clotho_mark_processed(entity_id: "<transcript_id>", process_name: "extraction")
   

Ontology Filtering

Before extracting, load the ontology for the relevant program(s):

clotho_get_ontology(entity_id: "<program_id>")

The ontology tells you what to look for AND what to ignore:

• Keywords — topics this program cares about. Prioritize signals matching these.
• Technical/social signals — specific signal types to watch for.
• Ignore list — topics to skip. If a signal matches an ignore entry, do NOT create an entity for it. Report it as skipped in the summary.

If a signal doesn't match any program's keywords or signal types, flag it for the user rather than auto-creating:

"I found a potential risk about 'supply chain delays' but it doesn't match any program's ontology. Should I create it, or is it noise?"

To add ignore rules:

clotho_update_ontology(entity_id: "<program_id>", add_ignore: "office logistics, catering, parking")

Search Before Create

MANDATORY: Before creating any entity, search for existing matches:

clotho_search(query: "<abstract theme>")

If a match exists, add an extracted_from relation to the existing entity instead of creating a duplicate. See the debrief-processor agent for the full search-before-create protocol.

Important Principles

• Human-in-the-loop: Present your extractions to the user for review before creating entities. Say what you found and ask for confirmation.
• Confidence: Be explicit about certainty. "I'm confident this is a decision" vs "This might be a risk, but it could also be a concern without action."
• Don't over-extract: Not every statement is a speech act. "Updates" (status reports) don't need entities — they're context.
• Respect the ignore list: If the ontology says to ignore a topic, ignore it. Don't second-guess.
• Do capture follow-ups: "Let's schedule a meeting about X" is an action item, not admin noise. Capture it.
• Watch the wrap-up: The end of a meeting is often the richest section for action items and next steps.
• Preserve provenance: Always create EXTRACTED_FROM relations back to the source transcript.