personal memory agent
Rewrite activity prompt to focus on observable changes
The activity generator was producing vague summaries describing what
was "visible" or being "reviewed" without evidence of actual activity.
Since the input is static screenshots, the prompt now explicitly:
- Requires identifying CHANGES between frames or SPEECH in audio
- Bans vague verbs (reviewing, monitoring, tracking, checking)
- Requires action verbs (wrote, sent, received, switched to)
- Adds pre-output check: "Can I point to a specific change?"
- Lists what to skip: unchanged windows, background apps
Tested on sample segments - outputs now show concrete actions like
"Modified Code: Added no_expire option" instead of "making headway
on the flightline repository".
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
+40
-33
+40
-33
muse/activity.md
+40
-33
muse/activity.md
···
1
1
{
2
2
3
3
"title": "Activity Synthesis",
4
-
"description": "Interprets each segment to extract meaning, intent, and searchability. Focuses on the 'why' behind actions - tasks, progress states, facets, and keywords for discovery.",
4
+
"description": "Synthesizes segment activity from screenshots and audio, focusing on observable changes and searchability.",
5
5
"color": "#00bcd4",
6
6
"schedule": "segment",
7
7
"output": "md"
···
12
12
13
13
# Segment Activity Synthesis
14
14
15
-
## Objective
15
+
## Core Rule
16
+
17
+
ONLY report what CHANGED between screenshots or was SPOKEN in audio.
18
+
If content looks the same across frames, skip it entirely.
19
+
20
+
## Your Inputs
21
+
22
+
- **Screenshots**: Sampled across this segment. Compare frames - what's different?
23
+
- **Audio**: Transcript of speech. What was said?
24
+
25
+
## Banned Language
26
+
27
+
Never use these words - they describe presence, not action:
28
+
- reviewing, monitoring, tracking, checking, observing, maintaining, managing
29
+
30
+
Use action verbs instead: wrote, sent, received, created, deleted, switched to, typed, said, discussed, decided
16
31
17
-
Interpret this segment to extract **meaning and searchability**. Your job is NOT to describe what happened (that's captured elsewhere) but to synthesize the underlying intent, progress, and context that makes this segment findable and understandable later.
32
+
## What to Report
18
33
19
-
## Core Focus
34
+
For each item, identify the CHANGE:
35
+
- "Typed message to X about Y" (text appeared)
36
+
- "Switched from Gmail to Terminal" (window focus changed)
37
+
- "Received reply from X" (new message appeared)
38
+
- "Said X about Y in meeting" (audio evidence)
20
39
21
-
### Tasks & Intent
22
-
What is the person really trying to accomplish? Look beyond literal actions to understand the underlying goals. Name the tasks concretely (e.g., "debugging the OAuth refresh flow" not "working on code").
40
+
If you cannot name the specific change, do not include it.
23
41
24
-
### Progress States
25
-
Characterize each task's working state:
26
-
- **Making headway**: Clear progress, solutions emerging
27
-
- **Blocked**: Stuck on errors, waiting, can't proceed
28
-
- **Exploring**: Investigating options, researching
29
-
- **Deciding**: Weighing alternatives, making choices
42
+
### Facets
43
+
Which project/context? Every segment has at least one.
30
44
31
-
### Facet Associations
32
-
**Required**: Identify which facet(s) this segment relates to. Every segment belongs to at least one facet. If work spans multiple facets, list all.
45
+
### Keywords
46
+
Project names, people, problems, decisions, tools.
33
47
34
-
### Searchable Keywords
35
-
What terms would someone use to find this segment later? Include:
36
-
- Project/feature names
37
-
- Problem descriptions ("authentication bug", "memory leak")
38
-
- Key decisions made
39
-
- People involved and their roles
40
-
- Tools or technologies central to the work
48
+
## Before Writing
41
49
42
-
## Output Format
50
+
For each item, ask:
51
+
- Can I point to a SPECIFIC CHANGE between screenshots?
52
+
- Or SPECIFIC WORDS spoken in audio?
43
53
44
-
Write **concise, scannable markdown**. Aim for brevity - a few paragraphs or bullet points, not a detailed narrative. Structure flexibly based on content:
54
+
If neither, omit it.
45
55
46
-
- Group by facet when multiple are present
47
-
- Note context switches and why they occurred
48
-
- Highlight blockers or breakthroughs
49
-
- Include emotional/energy state if notable (frustrated, energized, scattered)
56
+
## SKIP Entirely
50
57
51
-
## What NOT to Include
58
+
- Windows that look identical in first and last frame
59
+
- Apps open but showing same content throughout
60
+
- Background windows never brought to focus
61
+
- Anything you'd describe as "had open" or "was visible"
52
62
53
-
- Detailed chronological account of screen activity
54
-
- Exact commands, code excerpts, file contents
55
-
- Entity lists
56
-
- Time-by-time breakdown
63
+
## Output Format
57
64
58
-
Remember: This synthesis should make the segment **discoverable through search** and **quickly understandable** without reading the full transcript.
65
+
Concise markdown. Bullets preferred. Group by facet if multiple.