| name | structured-minutes |
| description | Turns recorded transcripts (of meetings, dialogues, seminars, talks) into comprehensive structured minutes. Consolidates spoken content by theme rather than time order, and mines small talk, tips, and anecdotes for insight. Use this skill whenever an audio transcript is provided and structured minutes are requested — for example, when the user asks to "make minutes", "structure this transcript", "summarize the meeting recording", "organize the seminar content", or "structure the dialogue notes". |
| license | MIT |
structured-minutes — Comprehensive structuring of minutes
A skill that turns spoken transcripts into theme-consolidated structured minutes.
The default assumption is that every utterance carries value, so insights also get mined from small talk, tangents, tips, and anecdotes.
1. Scope
- Internal meetings, external meetings, seminars, dialogues, interviews, study groups, workshops, etc.
- 30 minutes to 2 hours of speech per session (transcripts of roughly 10,000-50,000 characters)
- Roughly 2-10 speakers
Out of scope: when an organization has a domain-specific minutes format (medical chart notes, legal-consultation records, etc.), that domain skill takes precedence. This skill targets generic minutes structuring.
2. Input Contract
transcript:
text:
path:
title:
date:
speakers:
output_path:
include_timestamps:
3. Execution flow
Input (transcript)
│
▼
S1: Read the full text + extract elements (themes / facts / opinions / tips / small talk / emotions are all in scope)
│
▼ (user confirmation)
│
▼
S2: Generate structured minutes (consolidate by theme across time + create hierarchy)
│
▼
Output (.md file + minutes presented to the user)
After each stage, ask the user to confirm. On receiving feedback, revise the relevant stage and ask again.
4. S1: full read + element extraction
4.1 Goal
Extract every element from the transcript without omission and prepare the structuring blueprint.
4.2 Procedure
Step 1: read the full transcript
- Read the transcript from start to end
- Identify speakers (fallback to "Speaker A", "Speaker B" if names are missing, then confirm with the user)
Step 2: element extraction
Treat all of the following as extraction targets (small talk and tangents included; if an apparently off-topic utterance contains insight, pick it up):
| Category | Targets |
|---|
| Theme / keywords | Main theme, sub-themes, technical terms |
| Facts / numbers | Specific data, dates, monetary figures, KPIs |
| Opinions / judgments | Each speaker's view, agreements, open items |
| Issues / problems | Problems raised, risks |
| Proposals / solutions | Action items, improvement proposals |
| Anecdotes | Personal experience, examples, case studies |
| Tips / know-how | Practical techniques, tool tips |
| Small talk / tangents | Apparently off-topic utterances that nevertheless carry insight |
| Emotion / temperature | Strong agreement, concern, high-energy comments |
Step 3: theme-integration design
- Group utterances on the same theme regardless of time order
- Design the main-topic → sub-theme hierarchy
- Lay out the causal / logical flow within each group
Step 4: present the extraction
Present the following concisely and ask the user to confirm:
- The list of identified main topics
- The list of speakers (confirm names)
- Notes (sections where the information is unclear, audio was hard to make out, etc.)
→ Stop here and ask the user to confirm
5. S2: generate structured minutes
5.1 Goal
Integrate the S1 extraction and produce comprehensive structured minutes.
5.2 Core principles
- Comprehensiveness: structure every utterance (small talk / tips / anecdotes included)
- Value extraction: state the insight / lesson / applicability of each utterance
- Theme integration: consolidate the same theme regardless of time order
- Causality: lay things out so the logical flow is clear
- Adaptive hierarchy: use 2-5 levels of hierarchy depending on the content's complexity
5.3 Handling duplicates
- Merge: only when the content is fully identical and in the same context
- Keep as separate value: keep on a separate line when there is a subtle difference, a new angle, or an intentional emphasis
5.4 Output format
Output in Markdown. Filename: minutes-<theme or date>.md
# <meeting name / theme>
**Date**: YYYY-MM-DD
**Participants**: <speaker list>
---
### Summary of the proceedings
Summary of the entire minutes in structured prose. Bold the key parts.
---
## <main topic 1>
### <sub-theme>
- **<item>** [speaker][hh:mm:ss]
- <detail / anecdote / tip> [speaker][hh:mm:ss]
- <further detail> (for complex cases) [speaker][hh:mm:ss]
- <specific example / practical tip> (for very detailed cases) [speaker][hh:mm:ss]
### <sub-theme>
- **<item>** [speaker][hh:mm:ss]
- <content drawn from small talk / anecdote> [speaker][hh:mm:ss]
- **Insight**: lesson / learning from the above
- **Applicability**: where this could apply
- **Related**: interactions with other themes
(Repeat per theme)
5.5 Rules for timestamps and speaker names
- Extract only hours-minutes-seconds from the source's timestamps and record as
[hh:mm:ss]
- Append
[speaker][hh:mm:ss] at the end of every entry
- When the source has no timestamps, omit them (confirm with the user)
- Use the speaker names confirmed in S1 Step 4 ("Speaker A", "Speaker B" when anonymization is requested)
5.6 How to write insights
- Add "Insight: ..." per item as needed
- Add "Applicability: ..." when there is a practical application
- Add "Related: ..." when there is a link to another theme
- Not every item needs all three. Apply them to the high-value utterances
5.7 Quality criteria
- Bold the key parts
- Mark speculation / inference clearly
- Tag unclear information with
[needs confirmation]
- For specialized terminology, fix the wording at first appearance (e.g. AI / artificial intelligence, KPI / 重要業績評価指標) and stay consistent
6. Integration with compounding-wiki (optional)
This skill is usable standalone, but combining it with skills/knowledge-extract/SKILL.md enables the following flow:
1. structured-minutes converts a transcript → structured minutes (.md)
2. Place the .md into inbox/ (source_type: meeting_note)
3. knowledge-extract decomposes it in parallel into person / entity / theme
4. If a major claim needs verification, hand it to long-form-analysis
The structured-minutes output is raw minutes (a raw observation log) with no verification step. When a claim needs verification, hand it to long-form-analysis on Axis B.
7. Confidential handling
When the transcript contains any of the following, mask or exclude before transcribing into the structured minutes:
- Specific monetary figures / valuations / ratio numbers
- Personal contact info (email / phone / personal ID / chat account etc.)
- Specific compensation / individual HR-evaluation numbers
- Specific undisclosed-strategy timelines
- Comments touching the reputation of a third party where the minutes-value is low
Replace confidential sections with [REDACTED: <category>].
Do not store the raw transcript under a publicly accessible wiki tree (including inbox/ when integrated with compounding-wiki). When long-term retention is required, limit storage to an access-controlled location (local encrypted disk, access-controlled separate repo, dedicated storage, etc.). The structured minutes this skill produces must be composed only of content that is safe to share.
8. Common misunderstandings
Q1. "Does small talk really belong in minutes?"
Write it when it yields insight. Skip it when the analyst judges it as "just chatter". The criterion: "will the lesson still be there six months from now when this is re-read?"
Q2. "Does every utterance need a [hh:mm:ss]?"
When the source has timestamps, append them everywhere; this helps later when going back to the original. When the source has no timestamps, do not force them.
Q3. "What if speaker names are unknown?"
Fall back to "Speaker A", "Speaker B" and confirm with the user at S1 Step 4. Role labels such as "Moderator", "Presenter", "Questioner" are also acceptable.
Q4. "Should agreements and action items live in a separate section?"
For long minutes, a separate "Agreements" / "Action items" section at the tail is preferable. For short minutes (under 30 minutes), bolding inline is fine.
9. Related skills
License
MIT (this skill). See LICENSE at the repo root for details.
