today

Generate an actionable daily insight briefing — not a logbook.

Phase 1: Locate and Profile the Data

1. Resolve output directory (direct python mode):

OUTPUT_DIR="${OUTPUT_DIR:-$HOME/.claude/cobrain}"
echo "$OUTPUT_DIR"

2. Build today's file path and check existence:

OUTPUT_DIR="${OUTPUT_DIR:-$HOME/.claude/cobrain}"
TODAY_FILE="$OUTPUT_DIR/$(date +%Y%m%d)-raw.md"
test -f "$TODAY_FILE" && echo "exists" || echo "missing"

3. If the file does not exist or is empty, report:
   • "No cobrain entries captured today. The daemon may not be running — check with status."
   • Stop here.

Phase 2: Efficient Metadata Extraction (bash-first)

CRITICAL: Do NOT read the raw file yet. Use bash/grep to extract structured metadata first. This avoids wasting tokens on repetitive VLM output and <think> blocks.

4. Extract entry count, time range, and per-app frequency:

# Entry count and time range
echo "=== ENTRY COUNT ==="
grep -c '^### ' "$TODAY_FILE"
echo "=== FIRST ENTRY ==="
grep -m1 '^### ' "$TODAY_FILE"
echo "=== LAST ENTRY ==="
grep '^### ' "$TODAY_FILE" | tail -1
echo "=== APP FREQUENCY ==="
grep '^### ' "$TODAY_FILE" | sed 's/^### [0-9:]\+ · //' | sort | uniq -c | sort -rn

5. Identify distinct activity blocks — consecutive runs of the same app represent a single work session:

# Show app transitions (block boundaries) with timestamps
grep '^### ' "$TODAY_FILE" | awk -F ' · ' '{app=$2} app!=prev {print NR, $0; prev=app}'

This gives you the session structure: how many blocks, what sequence of activities, and approximate durations.

Phase 3: Intelligent Content Sampling

6. For each distinct activity block (consecutive same-app entries), read ONLY the first and last entry to understand what that session was about. This achieves full activity coverage at ~10% token cost.

   • Use grep -n to find line numbers of block boundaries
   • Use Read tool with offset/limit to read only those specific entries (typically 5-10 lines each)
   • For very short blocks (1-2 entries), reading the single entry is sufficient
   • Never read the entire file

7. Classify each activity block into one of these work categories:

   • Deep Work: focused productive tasks (coding, writing, spreadsheet analysis, design work)
   • Communication: work messaging (WeCom, Slack, email), meetings
   • Research/Browsing: web research, documentation reading, learning
   • Personal/Other: personal chat (WeChat non-work), system settings, activity monitor, idle time

Phase 4: Generate Insight Report

Produce a report with these four sections. Total output should be under 40 lines — dense and scannable. Write in a professional tone, like a personal executive assistant's daily brief.

Output Format

## Daily Brief — <date>

### Work Accomplished
- <concrete deliverable or task completed, in past tense>
- <another deliverable>
- ...
(Focus on WHAT was produced/achieved, not what apps were open)

### Time Allocation
- Deep Work: X hrs (XX%) — <primary activities>
- Communication: X hrs (XX%) — <work vs personal breakdown>
- Research: X hrs (XX%) — <topics>
- Other: X hrs (XX%)
(Total tracked: X hrs, from HH:MM to HH:MM)

### Workflow Observations
- <actionable pattern insight with specific numbers>
- <another observation>
(2-3 observations max. Focus on context-switching frequency, longest focus blocks, communication fragmentation, or late-night work patterns)

### Suggestions
- <1 concrete, constructive suggestion tied to today's data, with potential impact>
(1-2 suggestions max. Must reference specific numbers from today. Must have plausible economic or productivity value.)

---
**Headline:** <single sentence summarizing the day, suitable for a weekly digest>

Rules for Each Section

Work Accomplished:

• Derive from the CONTENT of sampled entries, not app names
• "Edited Q4 financial spreadsheet" not "Used wpsoffice"
• "Reviewed PR #142 and left feedback" not "Used Chrome for GitHub"
• If entry content is too vague to determine deliverables, note what was worked on at a category level

Time Allocation:

• Calculate durations from timestamps between block transitions
• Distinguish productive communication (WeCom work discussions, Slack) from personal/ambient (WeChat personal chat, social media)
• Round to nearest 15 minutes

Workflow Observations:

• Count actual app transitions from the block boundary data — each transition is a context switch
• Identify the longest uninterrupted work session (largest consecutive same-app block)
• Note if communication was batched (clustered) or scattered (spread across the day)
• Flag unusual patterns: late-night work, very short focus blocks (<10 min average), excessive context-switching

Suggestions:

• Must reference specific data from today (e.g., "your 47 WeCom transitions suggest...")
• Must propose a concrete change (e.g., "batch responses into 3 windows: morning, post-lunch, end-of-day")
• Must articulate the benefit (e.g., "could recover ~45 minutes of fragmented time")
• Never give generic advice like "take more breaks" without tying it to today's numbers