Most AI assistants are passive. A chief of staff wakes up.

A real chief of staff doesn't wait for prompts. They check your inbox at 8:15am. Your calendar at 8:30. Your CRM at 9. And only interrupt you when something actually needs you.

Every AI tutorial stops at "give it tools." That's where this one starts. The 8 files, 4 skills, and 3 cron jobs below define the missing operating system — who matters, what's urgent, when to act alone, when to escalate, and what to check every 15 minutes from 8am to 9pm.

Everything on this page is extracted verbatim from github.com/snarktank/clawchief — the public starter kit Ryan Carson released for his own chief-of-staff setup. The patterns below are the ones he runs in production.

The 3-layer rule Skills tell the agent how to act. Files tell the agent who to act for. Crons tell the agent when to wake up. Without all three, you have a chatbot. With all three, you have an operator.

Why separation matters

The entire value of the repo is the separation boundary between files. Each layer owns exactly one thing. Swap any one without breaking the others.

• Prioritization lives in priority-map.md. Nowhere else.
• Resolution policy lives in auto-resolver.md. Nowhere else.
• Meeting-note ingestion lives in meeting-notes.md. Nowhere else.
• Live task state lives in tasks.md. Nowhere else.
• Archive state lives in tasks-completed.md. Nowhere else.
• Local secrets live in TOOLS.md. Nowhere else.
• Recurring orchestration lives in HEARTBEAT.md plus the cron jobs. Nowhere else.

8 files. 4 skills. 3 crons. One operator.

Three layers. Each layer is portable. Each file does one thing.

Who is allowed to interrupt your day.

Without this file, the agent treats every email like an emergency. With it, only revenue-blocking signals from named people reach you.

The four urgency levels

The four action modes

• Interrupt now. Surface directly in your primary channel as a short, clear alert.
• Handle and summarize. The agent acts, then sends a one-line update if useful.
• Queue for digest. Hold for the next structured summary or heartbeat if still relevant.
• Ignore. No user-facing message needed.

How to use this file

When a new signal arrives from Slack, email, calendar, tasks, meeting notes, docs, or notes:

1. Map it to zero or more people.
2. Map it to zero or more programs.
3. Assign an urgency level.
4. Choose the action mode.

If a signal maps to no important people and no important programs, it should usually be ignored or batched and archived.

Default routing rules

These deterministic routes tell the agent which skill owns which class of signal:

• General inbox, scheduling, travel, calendar integrity, operational coordination → executive-assistant.
• Partner, referral, prospect pipeline, outreach tracker, lead verification, prospecting batches → business-development.
• Direct task CRUD, reprioritization, completion, current-task review → daily-task-manager.
• Morning due-date retuning and daily actionability prep → daily-task-prep.
• Meeting notes enter through executive-assistant, then route into business-development too if the note changes outreach or partner state.

Things that should usually be ignored

• Casual chatter with no action.
• Repeated notifications that add no new information.
• Documents or messages with no connection to a priority person or program.
• Speculative ideas without owner, deadline, or next step.
• Low-stakes activity already captured in the task list.

When to act alone vs. ask first.

The hardest problem with AI assistants: they either ask permission for everything or act on things they shouldn't. This file makes that decision deterministic.

After classifying a signal, the agent prefers to resolve the obvious next step instead of merely summarizing it. That means: do low-risk operational work in the same turn when authority is clear. Update the source of truth as part of doing the work. Interrupt the principal only when judgment, approval, ambiguity, or sensitivity requires it.

01Auto-resolve now

Low-risk, operationally clear, reversible or easy to audit. The agent does the work and updates the source of truth in the same turn.

02Draft and ask the principal

The next step is visible, but the principal should approve the wording or decision first. Legal, policy, investor, pricing, press, emotional.

03Escalate without acting

Too much ambiguity, risk, or missing authority to safely draft or send. One-line summary with the blocker and the recommendation.

04No action or archive

Noise, duplicative, already handled, or not worth surfacing. Marked read and gone.

The five-condition safe-auto-resolve checklist

The agent only acts alone when all five are true. If any condition fails, it falls back to draft-and-ask.

• The signal is clearly understood.
• The correct source of truth is known.
• The action is operational rather than strategic.
• Authority is already clear.
• A mistake would be low-cost and recoverable.

Safe auto-resolve lane

• Update tasks.md.
• Turn meeting-note action items into tasks.
• Add reminders or follow-up tasks.
• Update a tracker after a reply or booking.
• Send low-risk scheduling messages when authority is clear.
• Create, update, or cancel calendar events.
• Send a short factual follow-up on a known thread.
• Archive email after it's truly handled.

Draft-first lane

• Legal or policy answers.
• Investor, board, or fundraising messaging.
• Pricing-sensitive replies.
• Press, podcast, or speaking responses.
• Emotionally sensitive personal or family communication.
• Messages where wording matters more than logistics.

Escalate-without-acting lane

• Authority is unclear.
• The signal is contradictory or incomplete.
• There is reputational, legal, or financial risk without enough context.
• The action would expose private context the principal would not want disclosed.
• There is no reliable source of truth to ground the response.

Source-of-truth rule Do not auto-resolve from memory alone. Always ground the action in the relevant live source of truth before acting. Tasks → tasks.md. Outreach → CRM. Scheduling → live calendar. Priority → priority-map.md.

Good auto-resolve output examples

• "Booked it. Invite sent."
• "Updated the tracker and replied with the next step."
• "Turned the meeting-note follow-ups into tasks and handled the ones I could resolve."
• "Rescheduled for Tuesday at 2pm; nothing else needed from you."

The 11-step sequence the agent runs every 15 minutes.

This is the file that turns a passive chatbot into a proactive operator. Every cron firing, every wake-up, every check-in runs this exact sequence top to bottom.

1. Read the priority map.clawchief/priority-map.md
2. Read the auto-resolver policy.clawchief/auto-resolver.md
3. Read the meeting-notes policy and ledger.clawchief/meeting-notes.md plus workspace/memory/meeting-notes-state.json
4. Read the live task file.clawchief/tasks.md
5. Run the executive-assistant workflow for meeting-notes ingestion plus inbox, calendar, and scheduling triage.
6. If the signal is really about outreach tracker state, lead status, prospect pipeline, or partner follow-up — use the business-development workflow instead.
7. Auto-resolve low-risk operational items when the next step is obvious and authority is clear.
8. If meeting notes create principal tasks, add them to clawchief/tasks.md.
9. If the principal needs to know or act, send one short, direct update.
10. If there is no EA or biz-dev issue, use the remaining heartbeat value for at most one priority nudge anchored in a live program or open task.
11. If there is nothing useful to say, reply HEARTBEAT_OK.

The 8 governing rules

• Be proactive, but do not create noise.
• Keep this file as an orchestrator, not a duplicate workflow spec.
• Let the skills own their detailed procedures.
• Let priority-map.md own people and program priority.
• Let auto-resolver.md own auto-resolution policy.
• Let meeting-notes.md own meeting-notes ingestion policy.
• Let tasks.md own the live task state.
• Do not send repeated or near-identical nudges unless there is materially new information, a changed recommendation, or a concrete trigger.

The live task file and its overnight archive.

The agent reads tasks.md at the start of every heartbeat. Updates it in the same turn it changes anything. Treats it as the single source of truth across sessions, devices, and crons.

The 9 task-file rules (non-negotiable)

1. Read tasks.md before answering any question about current tasks or composing a check-in.
2. Treat it as the source of truth across all sessions.
3. When task state changes, update it in the same turn whenever practical.
4. When the assistant gets a task with a due date, add an assistant-owned task using the canonical due-date format.
5. When a task depends on an outside reply or future check-in, add a separate follow-up task with its own due date.
6. Scan for overdue and due-today assistant tasks before deciding what needs attention.
7. Keep long-term preferences in memory files, live state in tasks.md, and completed history in tasks-completed.md.
8. If a task change materially affects heartbeat behavior, ensure the heartbeat still points at tasks.md.
9. Use YYYY-MM-DD for all-day due dates and YYYY-MM-DD HH:MM TZ for timed due dates.

Turn every call into tasks automatically.

Most meeting notes die in a Google Doc nobody reopens. This file makes the agent treat notes as a live signal source — it reads them, extracts action items, and creates tasks before you even leave the meeting.

What the agent extracts from every note

• Explicit action items.
• Implied follow-ups.
• Deadlines and commitments.
• Promises the principal made.
• Introductions to send.
• Materials, links, or docs to share.
• Scheduling next steps.
• Outreach or partnership next steps.
• GTM or content tasks.
• Legal or policy questions that need review.
• Anything that should become a task, reminder, draft, or auto-resolve action.

The 7-step processing workflow

1. Search the configured source (Google Docs, AI notetaker output, shared folder) for recently shared or updated notes.
2. Compare candidate docs against workspace/memory/meeting-notes-state.json.
3. For each unprocessed note: read the note text, identify principal action items, assistant action items, follow-ups waiting on others, decisions and commitments, and auto-resolvable items.
4. Classify each extracted item through the priority map.
5. Run the auto-resolver policy.
6. Update tasks.md and any other live source of truth in the same turn.
7. Record the note as processed with timestamp and a short summary of what was extracted.

What counts as successful ingestion

A note is not handled just because it was read. It is only handled when:

• Tasks are added or updated in tasks.md.
• Follow-ups are created with their own due dates.
• Tracker, calendar, or inbox state is updated when appropriate.
• Auto-resolved actions are completed when safe.
• The processed-doc ledger is updated to prevent reprocessing.

Files are policy. Skills are procedure.

Each skill reads the relevant files, then runs a deterministic workflow. Skills live in ~/.openclaw/skills/ and are auto-loaded by the runtime.

Without crons, it's a chatbot. With them, it's an operator.

Three required jobs plus two optional ones. Each runs in its own session target. Each has a deliberately short prompt — the skill does the heavy lifting.

Channel adaptation

The repo is explicitly designed to run on whatever messaging surface you prefer. Keep workflows generic; only the delivery route changes.

Slack / Telegram / Signal / Discord / Google Chat

Use {{PRIMARY_UPDATE_CHANNEL}} and {{PRIMARY_UPDATE_TARGET}} in cron delivery blocks. Don't hardcode Slack unless you're explicitly Slack-only. Use one primary route — founder DM, private ops channel, or dedicated assistant thread. Don't spray updates across multiple surfaces.

Clone to first heartbeat in about 45 minutes.

Assumes OpenClaw is installed locally. GOG — the Google Workspace CLI — is a hard prerequisite. Set it up first.

Prerequisite — install GOG

1. Install GOG.brew install steipete/tap/gogcli
2. Create a Google OAuth desktop-client in Google Cloud Console.Enable: Gmail, Calendar, Sheets, Drive, Docs, People
3. Configure the OAuth consent screen and add the operating account as a test user.
4. Download the OAuth client credentials JSON.
5. Load credentials.gog auth credentials /path/to/client_secret.json
6. Add the operating account.gog auth add <email> --services gmail,calendar,drive,contacts,docs,sheets
7. Verify the account.gog auth list
8. Verify that Gmail message search, Calendar, and Sheets metadata all work.

The 13 install values to gather

Collect these before editing files. Every template placeholder maps to one of these.

{{OWNER_NAME}} · {{ASSISTANT_NAME}} · {{ASSISTANT_EMAIL}} · {{PRIMARY_WORK_EMAIL}} · {{PERSONAL_EMAIL}} · {{BUSINESS_NAME}} · {{BUSINESS_URL}} · {{TIMEZONE}} · {{PRIMARY_UPDATE_CHANNEL}} · {{PRIMARY_UPDATE_TARGET}} · {{GOOGLE_SHEET_ID}} · {{TARGET_MARKET}} · {{TARGET_GEOGRAPHY}}

The 7-step install flow

1. Clone the repo. Pull snarktank/clawchief into your OpenClaw workspace as a starter. Use it as a template, not a final config.
2. Install the 4 skills. Copy skills/executive-assistant, business-development, daily-task-manager, daily-task-prep into ~/.openclaw/skills/.
3. Install the workspace files. Copy clawchief/, workspace/HEARTBEAT.md, workspace/TOOLS.md, and workspace/memory/meeting-notes-state.json into ~/.openclaw/workspace/.
4. Replace all placeholders. Grep for {{ across the whole workspace. Replace each token with the real value from your 13-item list.
5. Customize priority-map + TOOLS. Strip the example people and programs. Add your real ones. Be specific about who escalates to P0. Be ruthless about what gets ignored.
6. Customize the BD outreach template. Edit skills/business-development/resources/partners.md. Change the subject and body to match your offer.
7. Create the 3 cron jobs. Copy cron/jobs.template.json into your runtime. Replace {{TIMEZONE}}, {{PRIMARY_UPDATE_CHANNEL}}, {{PRIMARY_UPDATE_TARGET}}. Enable the EA sweep first. Enable the two 2am crons after a week of tuning.

Common mistake Most people enable all 3 crons before customizing the priority map. You'll get 52 useless sweeps per day. Customize priority-map.md and TOOLS.md first, then enable the EA sweep, then the daily crons after a week of tuning.

Private files not shipped by the repo

The public starter kit intentionally does not ship these. Author your own if your setup depends on them:

AGENTS.md · SOUL.md · USER.md · IDENTITY.md · MEMORY.md · memory/

The install is good only if every item below passes.

This is the verbatim install-validation checklist from the repo. Walk it after the 7-step install.

GOG

• gog auth list shows the correct operating account.
• Gmail message search works.
• Calendar list and read works.
• Sheets metadata read works.
• Google Docs read works if meeting-notes ingestion is enabled.

Skills

• executive-assistant installed.
• business-development installed.
• daily-task-manager installed.
• daily-task-prep installed.

Workspace files

• priority-map.md installed.
• auto-resolver.md installed.
• meeting-notes.md installed.
• tasks.md installed.
• tasks-completed.md installed.
• HEARTBEAT.md installed.
• TOOLS.md installed.
• meeting-notes-state.json installed.
• All placeholders replaced.

Behavior verification

• Heartbeat reads source-of-truth files, not duplicated logic.
• Proactive updates route to the intended channel and target.
• Inbox sweeps use message-level Gmail search.
• Scheduling checks all relevant calendars before booking.
• Task system uses tasks.md as live source of truth.
• Prior-day completions archive to tasks-completed.md.
• Meeting notes are treated as a live signal source if enabled.
• BD work treats the outreach sheet as the live source of truth.
• Daily task prep promotes due-today items into ## Today.

Cron jobs

• Executive-assistant sweep exists.
• Daily task prep exists.
• Daily business-development sourcing exists.
• Optional nightly backup configured only if desired.
• Optional self-update enabled only if desired.

The 13 lessons baked into this repo.

These are the lessons the author learned from running this setup in production — documented in the repo's README and REVIEW-NOTES. Every one is a tested pattern, not a hypothesis.

• Keep a separate source-of-truth layer instead of burying everything in heartbeat text.
• Use Gmail message search, not thread-only search — unanswered outbound disappears from thread view.
• Check all relevant calendars before offering or booking time, including family and shared conflict calendars.
• Keep one canonical live task file and a separate completed-task archive.
• Separate task management (live CRUD) from daily task prep (morning restructure).
• Treat the outreach tracker as a live source of truth, not a reporting artifact. Update it in the same turn you act.
• Treat meeting notes as an operational signal source, not passive documents.
• Use a priority map to decide what matters — separate people, programs, and urgency.
• Use an auto-resolver policy to decide whether to act, draft, escalate, or ignore.
• Keep proactive update transport configurable instead of hardcoding one surface.
• Keep cron prompts DRY and let skills hold the workflow logic.
• Keep optional backup and self-update as explicit choices, not assumptions.
• When a task depends on an outside reply, add a follow-up task before ending the turn — don't rely on inbox memory.

The math — what this actually substitutes for.

Comparing apples to apples. The named tools and roles a fractional founder typically pays for to get the same coverage.

This is not AI replaces humans. This is: the boring 80% of an EA's day is now files and crons. The strategic 20% — judgment, taste, relationships — is still yours.