Skip to content

Commit ad0172f

Browse files
darthwalshdarthwalsh
committed
notes
1 parent 8ec33fa commit ad0172f

1 file changed

Lines changed: 147 additions & 0 deletions

File tree

‎apps/AIInMeetings.md‎

Lines changed: 147 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -75,3 +75,150 @@ Real-time collaborative ChatGPT sessions (Google-Docs-style) **not supported** i
7575
- Screen share + driver model
7676
- Slack thread + ChatGPT app
7777
- Shared prompt doc + everyone runs independently
78+
79+
## Async: Structuring a Transcript
80+
81+
After transcribing (see [[SpeechRecognition]]), two useful post-processing steps:
82+
83+
**1. Polish the raw transcript** — clean up whisper hallucinations, filler words, merge short segments:
84+
85+
---
86+
87+
You are given a transcript file split for a phone call with Me and You speakers. Produce a **polished transcript** (no preamble/commentary).
88+
89+
**Requirements:**
90+
* If the same phrase is repeated 5+ times (whisper model hallucination), replace with `[Transcription Failure]`
91+
* If you see `You know, You know, You know, You know, You know, You know, You know,` → `You know... [Transcription Failure]`
92+
* Detect transcription failures **before** normal repetition condensing.
93+
* Keep the general text formatting like `[00:12] Me: I said something.`
94+
* Merge short speech sections into natural sentence boundaries (speaker change, long pause, or output line getting long)
95+
* Infer misheard words **only when there is a clear, unambiguous technical term implied by immediate context**
96+
* "cash server" → "cache server"
97+
98+
99+
Polished written prose:
100+
- **Remove filler words**, false starts, and verbal tics unless they materially affect meaning.
101+
- **Resolve self-corrections** into the final intended phrasing.
102+
- **Lightly normalize grammar and sentence structure** (spoken fragments → complete sentences where appropriate), inferring commas but avoiding em-dashes.
103+
- **Condense repetitions** that do not add new information.
104+
- Preserve **speaker turns and timestamps**, but allow timestamps to represent logical breaks rather than exact utterance timing.
105+
- Retain **interruptions or overlaps only when they affect meaning**, marking them concisely (e.g., `[interrupting]`).
106+
107+
Do **not** introduce new information, interpretations, or summaries.
108+
Only include a postscript if there was `[Transcription Failure]`, with the timestamps.
109+
110+
---
111+
112+
**2. Diagram the conversation** — after polishing, extract structure (chapters + outline + decisions):
113+
- [ ] Try this
114+
Paste the polished transcript into an AI with this prompt:
115+
116+
```text
117+
You are given a many-minute conversation transcript. The transcript includes:
118+
- speaker labels (e.g., "Carl:", "Alex:")
119+
- timestamps (e.g., "[07:32]" or "07:32")
120+
121+
Your job: produce a structured "conversation diagram" that starts with CHAPTERS (time-ranged segments), then for EACH chapter outputs:
122+
1) Outline (hierarchical bullets)
123+
2) Open Questions
124+
3) Tasks / Action Items
125+
126+
Hard requirements:
127+
- Start with a short "Chapter List" that has 4–8 chapters max for a 20-minute transcript.
128+
- Every chapter MUST include an explicit time range like "07:10–10:45".
129+
- Prefer concrete paraphrases over long quotes. Only quote short phrases when necessary.
130+
- Tie claims to evidence by including timestamps (and speaker names when relevant).
131+
- Be conservative: if something is implied but not stated, mark it as "inferred" and include the evidence timestamp.
132+
133+
Output format (follow exactly):
134+
135+
# Chapter List
136+
1. <Chapter title> (<start–end>)
137+
2. ...
138+
139+
# Chapter 1: <Title> (<start–end>)
140+
## Outline
141+
- <Top-level point> [timestamp]
142+
- <Subpoint> (Speaker) [timestamp]
143+
- <Subpoint> (Speaker) [timestamp]
144+
- <Top-level point> [timestamp]
145+
146+
## Open Questions
147+
- Q: <question text> [timestamp where it came up]
148+
- Context: <why it's open / what was said> (Speaker) [timestamp]
149+
- Owner (if stated): <name or "unspecified">
150+
- Needed to answer: <what info/decision is missing>
151+
152+
## Tasks / Action Items
153+
- [ ] <task verb + object> — Owner: <name or "unspecified"> — Due: <date or "unspecified"> [timestamp]
154+
- Acceptance criteria: <what "done" looks like, if stated or inferred>
155+
- Dependencies/Risks: <if mentioned>
156+
157+
After all chapters, add:
158+
# Cross-Chapter Summary
159+
## Decisions Made
160+
- <decision> [timestamp]
161+
## Key Themes
162+
- <theme> [chapter refs]
163+
## Parking Lot (Deferred)
164+
- <deferred topic/question> [timestamp]
165+
## Task Rollup
166+
- <task> — Owner — Due [timestamp]
167+
168+
Now do the work on the transcript below.
169+
170+
----------------------------------------------------------------
171+
EXAMPLES (these are examples of the desired style; do NOT treat them as facts)
172+
173+
Example Chapter List:
174+
1. Framing the problem (00:00–03:20)
175+
2. Constraints and requirements (03:20–07:05)
176+
3. Options discussed (07:05–12:40)
177+
4. Converging on a plan (12:40–17:30)
178+
5. Next steps and owners (17:30–20:00)
179+
180+
Example Chapter (style example):
181+
# Chapter 2: Constraints and requirements (03:20–07:05)
182+
## Outline
183+
- Define non-negotiables for the solution [03:25]
184+
- "Must work offline for installers" (Alex) [03:40]
185+
- "Avoid new auth prompts during silent install" (Carl) [04:05]
186+
- Identify integration constraints [05:10]
187+
- Existing dependency on <System X> (Alex) [05:22]
188+
- Concern about rate limits (Carl) [06:10]
189+
190+
## Open Questions
191+
- Q: What's the acceptable timeout threshold during install? [06:35]
192+
- Context: Carl raised concern about retries; no threshold agreed. (Carl) [06:35]
193+
- Owner (if stated): unspecified
194+
- Needed to answer: installer performance budget; historical telemetry
195+
196+
## Tasks / Action Items
197+
- [ ] Pull installer telemetry for average install duration — Owner: Alex — Due: unspecified [06:50]
198+
- Acceptance criteria: report with p50/p90/p99 and notes about slow-path causes
199+
- Dependencies/Risks: access to telemetry dashboard
200+
```
201+
202+
- Good for: decision meetings, troubleshooting calls, negotiation
203+
204+
### Topic timeline ("chapters")
205+
206+
Shows when topics start/stop, pacing, and digressions — already captured by the chapter list above. Can also be rendered visually:
207+
208+
- Mermaid timeline (Markdown) in Obsidian/GitHub
209+
- Miro / FigJam (drag blocks along a horizontal line)
210+
211+
Example:
212+
```
213+
00:00–03:20: Problem framing
214+
03:20–09:10: Constraints + requirements
215+
09:10–14:30: Options
216+
14:30–18:40: Decision
217+
18:40–20:00: Next steps
218+
```
219+
220+
### Entity/Concept map (future research)
221+
222+
Shows the mental model that emerged from the conversation. Nodes = concepts/products/teams/risks; edges = "relates to / blocks / depends on". Edges can be annotated with timestamps of where they came up.
223+
224+
Tools: Obsidian Canvas, Excalidraw, yEd, draw.io, Miro mind map

0 commit comments

Comments
 (0)