personal memory agent
improve Gemini transcription prompt for consistent JSON output
Remove markdown code block and ellipsis from output format spec, replace
with inline JSON example and explicit "do not wrap in an array" constraint.
Gemini was intermittently returning non-standard shapes ([{"segments": [...]}],
bare arrays) — clearer instructions produce consistent {"segments": [...]}
responses.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
+7
-11
+7
-11
observe/transcribe/gemini.md
+7
-11
observe/transcribe/gemini.md
···
15
15
16
16
## Output Format
17
17
18
-
Return JSON only:
18
+
Return a single JSON object with one key, "segments", containing an array of statement objects. Do not wrap the object in an array.
19
19
20
-
```json
21
-
{
22
-
"segments": [
23
-
{"start": "01:23", "speaker": "Speaker 1", "text": "First statement in clip"},
24
-
{"start": "01:28", "speaker": "Speaker 2", "text": "Response from another person"},
25
-
{"start": "01:35", "speaker": "Speaker 1", "text": "Next statement"},
26
-
...
27
-
]
28
-
}
29
-
```
20
+
{"segments": [{"start": "01:23", "speaker": "Speaker 1", "text": "First statement in clip"}, {"start": "01:28", "speaker": "Speaker 2", "text": "Response from another person"}]}
21
+
22
+
Each statement object has exactly three fields:
23
+
- "start": absolute MM:SS timestamp in the original recording
24
+
- "speaker": consistent speaker label like "Speaker 1"
25
+
- "text": verbatim transcription of what was said
30
26
31
27
## Guidelines
32
28