personal memory agent
Simplify segmentation prompt with mechanical rules for flash model
Replace abstract topic-detection instructions with concrete steps a
flash model can follow: divide total lines by 100 for segment count,
space evenly, adjust to nearest speaker change. Tested against 23-page
timestamp-free meeting transcript — produces 14 well-distributed segments.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
+17
-26
+17
-26
think/detect_transcript_segment.md
+17
-26
think/detect_transcript_segment.md
···
4
4
label: Segmentation
5
5
group: Import
6
6
---
7
-
You are a transcript analyzer that splits transcripts into ~5-minute segments.
7
+
Split a transcript into ~5-minute segments.
8
8
9
-
TASK: Find segment boundaries and return their line numbers with absolute time-of-day timestamps.
9
+
INPUT:
10
+
- First line: "START_TIME: HH:MM:SS"
11
+
- Remaining lines: numbered "N: content"
10
12
11
-
INPUT FORMAT:
12
-
- First line: "START_TIME: HH:MM:SS" - the absolute start time of this transcript
13
-
- Remaining lines: Transcript with line numbers prepended as "N: content"
13
+
OUTPUT: JSON array of {"start_at": "HH:MM:SS", "line": N}
14
14
15
-
OUTPUT FORMAT:
16
-
- JSON array of objects with "start_at" and "line" fields
17
-
- "start_at": Absolute time-of-day in HH:MM:SS format
18
-
- "line": Line number where this segment begins
19
-
- Example: [{"start_at":"12:00:00","line":1},{"start_at":"12:05:23","line":42}]
15
+
RULES:
16
+
1. First segment is always {"start_at": START_TIME, "line": 1}
17
+
2. If the transcript has timestamps, use them to find ~5-minute boundaries. Add relative timestamps (00:05:30) to START_TIME to get absolute times.
18
+
3. If the transcript has NO timestamps, follow these steps:
19
+
a. Count the total lines in the transcript
20
+
b. Divide by 100 to get the number of segments (round up, minimum 2)
21
+
c. Space segments roughly evenly by line count
22
+
d. Adjust each boundary to the nearest topic or speaker change
23
+
e. Space the "start_at" times evenly across 5 minutes per segment from START_TIME
24
+
4. All "start_at" times must be absolute HH:MM:SS
25
+
5. Do NOT put boundaries in the middle of someone speaking
20
26
21
-
SEGMENTATION MODES:
22
-
23
-
1. **Timestamped transcripts** — if the text contains timestamps (relative like 00:05:30 or absolute like 14:30:22), use them to find boundaries near 5-minute intervals. Convert relative timestamps by adding to START_TIME.
24
-
25
-
2. **Timestamp-free transcripts** — if the text has NO timestamps (e.g. just speaker labels and dialogue), segment by **topic and conversation shifts** instead:
26
-
- Find natural break points where the conversation changes subject
27
-
- Estimate time from position: assume ~130 words/minute speaking rate, calculate total duration from word count, then assign proportional timestamps from START_TIME
28
-
- Aim for segments roughly 5 minutes of estimated speaking time, but prioritize clean topic breaks over exact intervals
29
-
- NEVER distribute lines uniformly — segments should vary in size based on where topics actually change
30
-
31
-
REQUIREMENTS:
32
-
1. First segment always starts at START_TIME on line 1
33
-
2. All output times must be absolute HH:MM:SS
34
-
3. Every transcript gets multiple segments unless it is extremely short (under ~2 minutes estimated)
35
-
36
-
RESPONSE: Return only the JSON array, no additional text.
27
+
Return only the JSON array.