Chief Of Staff

Chief of Staff

A lean daily briefing scannable in 30 seconds. Focuses on what matters: schedule, week priorities, and yesterday's agent work.

Morning briefing workflow

Step 1: Load context

Dispatch these sources as parallel subagents—one per source-group, and one per Google account for email—rather than loading them sequentially. They are almost entirely independent, so fanning out cuts the briefing's wall-clock time substantially. Delegate the pure lookups (calendar, day tracker, Todoist, Claude Code summaries, raw email/WhatsApp fetches) to Haiku subagents. Keep the judgement calls in the main loop: which emails genuinely need a reply, which WhatsApp messages clear the bar, and the final assembly. Attendee context is one chain per attendee (resolve person → read project files), but attendees are independent of each other, so they fan out too.

Gather information from these sources:

Monday context: On Mondays, also look for the previous week's review at work/week-reviews/2026-W{XX-1}-review.md. If it exists, read it and use it to enrich the "Week priorities" section (mark items as done/missed based on the review) and the "What can I help with?" suggestions (e.g. carry forward unfinished work, flag patterns from the review). If no review exists, note that in the briefing and suggest running /week-review first.

Calendar attendee context: For each named attendee of today's events (excluding Peter), look them up in the people registry at ~/.agents/data/people.json. The key convention is <first>-<last> lowercased and hyphenated (e.g. "Jane Smith" → jane-smith).

If the attendee's entry has a default_project value (e.g. ai-wow, coaching-client, acme-corp):

1. Resolve the project name to a folder path via ~/.agents/references/project-map.md.
2. Read both CLAUDE.md and AGENTS.md from that project folder (either or both may exist).
3. If those files name a per-person notes file in the project's context/ folder (e.g. context/ph-js-weekly-calls.md), read that too.

Use the loaded context to write the "Notes" column for that event — what the standing relationship is, what was last discussed, what's pending. This is what catches things like "Jane Smith = AI Wow peer mentor, weekly check-in" without guessing.

If the attendee is not in the registry, write "No project context — unregistered" in the Notes column rather than inferring. If it looks like a recurring relationship, suggest registering them via /summarise-granola (which has a "register unregistered people" step).

Todoist overdue check: Every morning briefing should check td for incomplete tasks due in the previous 7 days. Mention them explicitly somewhere in the briefing (usually in Week priorities, Yesterday, or a short Overdue tasks subsection if there are several).

Suggested approach:

# Get tasks due in the last 7 days that are still incomplete.
# Note: `td task list` only returns active (non-completed) tasks, so no !completed clause is needed
# — including it returns HTTP 400 because Todoist's filter syntax doesn't support it.
td task list --filter "due after: -8 days & due before: today" --json

If that returns nothing, there are no recent overdue tasks—omit the section. For each overdue task, include content, due date, and project (if available). If there are many, group by project.

Email triage: Search all three accounts for emails that likely need a timely reply. Run these searches:

# Personal account — unread inbox, last 3 days
gog --json gmail search 'label:inbox is:unread newer_than:3d' --account your-email@gmail.com --max 30

# T3A account — unread inbox, last 3 days
gog --json gmail search 'label:inbox is:unread newer_than:3d' --account your-email@example.com --max 20

# IWR account — unread inbox, last 3 days
gog --json gmail search 'label:inbox is:unread newer_than:3d' --account your-product@gmail.com --max 20

From the results, select emails that are actually important and likely need a reply. Use judgement, but always include:

• Any TYPE III AUDIO client (check sender against known clients)
• Emails from current advisory-client domains
• Emails from your accountant
• Personal emails from real humans that look time-sensitive

If data/triage-allowlist.md exists, read it for the specific high-priority senders to always include (client domains, accountant addresses, etc.). This file is local-only and never published.

Always exclude:

• Todoist task notification emails
• Newsletters, marketing, automated notifications
• Receipts, shipping updates, and other transactional noise

Prefer false positives over false negatives. Peter will give feedback to reduce noise over time.

If no important emails are found, omit the section entirely.

No Slack mentions. Peter checks those at lunchtime. Urgent items reach him via Signal/push.

WhatsApp (last 24h): Query the bridge DB on mini for messages in the last 24 hours. Full schema and query patterns: /Users/ph/.agents/references/whatsapp-bridge.md. Use the reference's standard queries with FROM = 24h ago, which already filter out noisy groups (>3 distinct senders) and empty-content rows.

Scan for anything that genuinely deserves Peter's attention today: questions awaiting a reply, plans being made for today/this week, emotional or LTP-related threads, commitments to follow up on. Be highly selective — most WhatsApp volume is chit-chat or logistics that doesn't need surfacing. Include at most a handful of bullets under a ## WhatsApp section, or omit entirely if nothing rises to the bar. Err on the side of omission for the daily briefing (unlike email, where false positives are OK).

Never send WhatsApp messages. Read-only.

Week number calculation: Use ISO week format (2026-WXX).

Day tracker summary generation:

Before reading day-tracker data, ensure yesterday's summary exists:

1. Check if yesterday's JSON has a summary object: cat /Users/ph/Documents/day-tracker/data/daily/YYYY-MM-DD.json | python3 -c "import json,sys; print('Has summary:', 'summary' in json.load(sys.stdin))"
2. If missing, generate it:

   cd /Users/ph/.claude/skills/day-tracker && python3 cli.py summary YYYY-MM-DD
   

3. Then read the summary from the JSON file

Claude Code + Codex digest:

Read from the day-tracker completions sidecar file:

cat /Users/ph/Documents/day-tracker/data/daily/YYYY-MM-DD.completions.json

If the file doesn't exist, run the rollup first:

cd /Users/ph/.agents/skills/day-tracker && python3 scripts/daily-rollup.py YYYY-MM-DD

The agent_sessions key contains session counts grouped by project. Other keys (git_commits, emails_sent, google_docs_edited, calendar_events) provide additional context for the briefing.

Step 2: Present briefing

Display the briefing in chat, save it to data/briefings/YYYY-MM-DD.md, update Day.md with today's schedule via update-day-schedule.py (see Fetching calendar events), and open the saved briefing in Typora:

open -a Typora /Users/ph/.agents/skills/chief-of-staff/data/briefings/YYYY-MM-DD.md

Structure the briefing as follows:

# Good morning

**[Day] [Date] [Month] [Year] · W[XX]**

> **Mantra:** [Week mantra from the week plan, if set]

## Today's schedule

| Time | Event | Notes |
|------|-------|-------|
| HH:MM | Event name | Contextual insight from week plan / project knowledge |

[If no events before afternoon: "No morning commitments—deep work day."]

## Needs reply

| Sender | Subject | Age | Account | Context |
|--------|---------|-----|---------|---------|
| [Name] | [Subject line] | [e.g. 2d] | [personal / t3a / iwr] | [One-line context: why this matters or what action is needed] |

[Omit this section entirely if no important emails were found.]

## Week priorities

| Priority | Status | Notes |
|----------|--------|-------|
| [Key result from week plan] | [Done / X/Y done / In progress / Not started] | [Concrete progress indicator, what's next] |

## Yesterday

| Project | Sessions | What happened |
|---------|----------|---------------|
| [project-name] | N | One-line summary of work done |

[Day tracker: Xh tracked—breakdown by category]
[Or: "No day tracker data for yesterday."]

## What can I help with?

1. **[Action]** — grounded in today's situation (calendar, week plan, yesterday's unfinished work)
2. **[Action]**
3. **[Action]**
4. **[Action]**
5. **[Action]**

Key rules for content:

• Needs reply is curated from the Gmail search results. Only include emails that genuinely look like they need a human reply. Err on the side of inclusion—Peter will give feedback to tighten the filter.
• Schedule notes must be contextual and insightful, drawing on week plan and project knowledge—not just restating calendar entries.
• Week priorities are extracted from the current week plan's key results, with concrete progress indicators (e.g. "2/5 done", "Drafted, needs review").
• Yesterday comes from the Claude Code + Codex digest. Concise: project, session count (noting source if mixed), one-line summary.
• "What can I help with?" must be grounded in today's actual situation. Each option should be something Claude can actually do right now. Not generic.

Day note update (standalone)

Triggered by /cos schedule, /gm schedule, or "update day note".

Use case: Refresh today's schedule in Day.md mid-day.

Steps:

1. Fetch events from both calendars (see Fetching calendar events)
2. Update Day.md via update-day-schedule.py
3. Print today's schedule in the chat in HHMM Event name format
4. Confirm: "Day.md updated — N events written"

No briefing file saved, no digest generated, no week priorities or yesterday section.

Day recap

Triggered by "day recap", "what done", "what have i done", "what did i do", or /recap.

Use case: Mid-day or end-of-day accountability check. Compare plan vs actual work, flag distractions. Output should be screenshot-friendly (Peter often screenshots it to his accountability partner).

Step 1: Load context

Fan these out as parallel subagents—agent sessions, git commits, active directories, calendar, day tracker, and Mochi are all independent lookups. Delegate them to Haiku and assemble once they return; keep the plan-vs-actual judgement in the main loop.

Gather from these sources:

Session discovery: Use generate_digest.py --json for agent sessions only. For git commits, call the day-tracker collector directly (see table above). These are separate systems.

Step 2: Present recap

Display in chat only—no file saved.

# Day recap

**[Day] [Date] [Month] · [HH:MM]**

## What you did

| Project | Work | Source |
|---------|------|--------|
| [project] | One-line summary | N sessions, N commits |

[If meetings happened: list them briefly]
[Day tracker: Xh tracked — breakdown by category. Or: "No day tracker data yet."]

## Plan vs actual

**Today's plan:** [Copy the day-by-day line from week plan]

| Planned | Status |
|---------|--------|
| [planned item] | Done / In progress / Not started / Skipped |

## Off-plan check

**"NOT doing" list:**
| Item | Verdict |
|------|---------|
| [item from week plan "NOT doing" section] | Clean / Violated — [evidence] |

**Commitments to accountability partner:**
| Commitment | Verdict |
|------------|---------|
| [commitment from week plan] | Kept / Broken — [evidence] |

## Verdict

[One sentence: on track, drifting, or off-plan. Be direct.]

## Appendix: git commits

[List all git commits from today across all repos, grouped by repo. Use compact format:]

| Time | Repo | Message |
|------|------|---------|
| HH:MM | repo-name | Commit message (first line only) |

Key rules:

• Be honest and direct. The point is accountability, not encouragement. If Peter spent 3 hours on something not in the plan, say so plainly.
• Evidence-based verdicts. Every "Violated" or "Broken" must cite specific agent sessions, commits, or active directories as evidence.
• Screenshot-friendly. Keep it compact. No verbose descriptions. Tables over prose.
• "NOT doing" is critical. The week plan's "What I'm NOT doing this week" section lists specific distractions Peter has committed to avoiding. Check agent sessions and active directories against this list explicitly.
• Commitments to accountability partner. The "Commitments (meta)" section in the week plan contains promises like "No podcatcher Mon–Wed". Check these against the evidence.
• Mochi commitment. If the week plan includes a Mochi commitment, check mochi reviewed --json to verify. Do NOT rely on day-tracker screenshots—Mochi reviews happen on the phone app or via Telegram/Hermes.
• Time of day matters. At noon, "Not started" on an afternoon task is fine. At EOD, it's a miss. Calibrate accordingly.

Data paths reference

/Users/ph/Documents/Projects/
├── plans-and-reviews/
│   └── work/
│       └── week-plans/2026-WXX-plan.md
└── [project folders]/CLAUDE.md  # plus AGENTS.md, context/

/Users/ph/.agents/data/
└── people.json  # calendar attendee → default project

/Users/ph/Documents/day-tracker/data/
└── daily/YYYY-MM-DD.json

/Users/ph/.agents/skills/chief-of-staff/
└── data/
    ├── briefings/YYYY-MM-DD.md
    ├── claude-code-summaries/YYYY-MM-DD.json
    └── overnight-results/

Integration with existing skills

Weekly rhythm

Fetching calendar events

Fetch from both calendars:

gog cal events primary --from today --to today --account your-email@gmail.com
gog cal events your-meetings-calendar-id@group.calendar.google.com --from today --to today --account your-email@gmail.com

The Meetings calendar returns times in UTC—add 1h for CET (or 2h for CEST). Merge both calendars, sort by time, format each event as HHMM Event name.

Update Day.md by passing each event as an argument:

python3 /Users/ph/.agents/skills/chief-of-staff/scripts/update-day-schedule.py "0800 Standup" "1400 Call with Alice"

No arguments → inserts "No events scheduled."