onewave-ai/meeting-to-tasks

Meeting to Tasks

Transform raw meeting transcripts into structured, actionable outputs: decisions, action items, owners, deadlines, and open questions -- including implicit commitments participants may not realize they made.

Contents

• references/extraction-schema.md -- YAML field schemas for every extraction phase
• references/markers.md -- Decision/action/implicit markers, priority + deadline inference, confidence scoring
• references/output-templates.md -- Meeting summary, task files (Linear/GitHub/generic), follow-up email, directory layout
• references/meeting-types.md -- Meeting-type detection and per-type focus

Workflow

1. Accept input. Handle plain text, Markdown, pasted text, audio transcript exports (Otter.ai, Fireflies, Rev, Zoom, Google Meet, Microsoft Teams), or structured notes.
2. Validate. Confirm there is enough content (~100 words minimum), identifiable speakers, a discernible topic, and any timestamps. If input is too sparse, ask for the full transcript, the participant list, or the meeting purpose before proceeding.
3. Read the full transcript first. Do not extract from a partial read -- later context can change earlier interpretation.
4. Detect the meeting type and adjust focus per references/meeting-types.md.
5. Extract in phases using the schemas in references/extraction-schema.md: metadata, decisions, action items, open questions, parking lot, discussion points. Apply the markers and inference rules in references/markers.md. Preserve the exact source quote for every item.
6. Identify speakers. Build the participant list from speaker labels and track who said what. When labels are absent, infer from context clues, otherwise mark "Unassigned" and flag for the user to assign.
7. Score confidence on every item (high/medium/low). Flag all low-confidence items for human review per references/markers.md.
8. Resolve conflicts and ambiguity: flag contradictory commitments as "CONFLICT" with both versions; list all candidates for unclear ownership; translate vague deadlines ("soon", "ASAP") to specific dates and flag for confirmation; offer both readings of ambiguous scope; consolidate duplicates.
9. Generate outputs using references/output-templates.md. Confirm which project management format(s) the user wants before generating task files. Always generate the follow-up email, even if unasked.
10. Post-process: deduplicate overlapping items; map dependencies; highlight the critical path; report owner imbalance; warn on deadline clustering; list items missing owners or deadlines.

Rules

• Make action items actionable: start each with a verb, specific enough that the owner knows exactly what to do. Transform "Think about pricing" into "Research competitor pricing and propose new pricing tiers by [date]."
• Err toward capturing too much; a flagged low-confidence item beats a missed commitment.
• Do not editorialize. Report what was said; keep any analysis in clearly labeled "Analysis" sections.
• Respect privacy. If the meeting contains HR issues, personal matters, or confidential information, flag it and ask how to handle it in outputs.
• Handle multi-language transcripts: extract in the primary language and note language-specific nuances.
• Maintain an extraction log noting confidence levels, ambiguities, and decisions made during extraction.

Quick Commands

• "Extract from [file]" -- Full extraction from the specified transcript
• "Just action items" -- Action items only (skip decisions, questions, summary)
• "Generate tasks for Linear" / "Generate tasks for GitHub" -- Action items in that format
• "Draft follow-up email" -- Follow-up email only
• "Who owes what?" -- Owner-grouped view of all action items
• "What was decided?" -- Decisions only
• "What is still open?" -- Open questions and unresolved items