Scheduled Check-Ins

Exposing Analysis Over MCP covers asking your assistant a question whenever one occurs to you. This page covers the other shape: an agent that reads your data on a schedule and writes to you first — a morning brief, a Sunday review — without you having to remember to ask.

Like By Automation, this isn't a feature you switch on. It's a recipe, and everything that makes it feel like yours is a line in a prompt you control: the personality, what it looks at, how often it runs, and where it lands. What follows is a starting point, not a template you're meant to keep.

Setting up a Claude Routine

A Claude Routine runs on Anthropic's infrastructure on a schedule you choose. Anthropic's Routines documentation is the reference; these are the parts specific to Cohearence.

1. Connect Cohearence as a connector on your Claude account at claude.ai/customize/connectors — see Signing In From Your Client. This is the step that trips people up. A routine can only use account-level connectors. If you added Cohearence locally in the Claude Code CLI (with claude mcp add or a .mcp.json file), it will not appear in the routine's connector list, and your runs will quietly have no access to your data.
2. Create the routine at claude.ai/code/routines with New routine, in Claude Code Desktop under Routines → New routine, or by running /schedule in the CLI.
3. Choose Cloud, not Local. A local task only runs while your computer is awake and the app is open — miss that window and the run is skipped, or fires late against stale data. A cloud routine runs whether or not your machine is on, which is the entire point of a brief waiting for you in the morning.
4. Paste the prompt below as the routine's instructions and pick a schedule. The minimum interval is one hour; daily is the natural cadence for a check-in.
5. Under Connectors, keep Cohearence and remove any the routine doesn't need. During a run the agent can call every tool on an included connector without stopping to ask.
6. Hit Run now once and read what comes back before you trust the schedule.

Each run becomes a session you can open — and reply to, which is what turns a brief into a conversation. Ask a follow-up and the agent digs into your data right there. Worth knowing: a green status in the run list only means the session finished without crashing, not that it worked, so open the first few and confirm it actually reached your data. Routines are a research-preview feature on paid Claude plans, so the details above may shift.

Other runners work the same way — a Cursor Cloud Agent, an OpenClaw instance, any cron host you keep running. The difference is that you configure the MCP connection yourself, and if there's no browser to sign in with, that's the case for a Personal Access Token.

A prompt to start from

Fill in the bracketed parts and delete anything that doesn't apply to you.

You are my daily health correspondent — a sharp, warm observer with access to my
health and journaling data through the Cohearence MCP server. Each day, read what
actually happened, connect it to what I logged, and give me a short, honest brief.

GATHER (in this order — do not skip ahead):
1. list_hunches — relationships I've explicitly said I care about. These are the
   spine of the brief.
2. catalog — note each metric's tracking_cadence (my declared intent) and which
   metrics are text-type. Do NOT rank by activity count: synced wearable metrics
   always dominate activity counts, and my hand-logged metrics are the
   interesting half.
3. query_measurements — raw data for the last 30 days across every metric,
   including text metrics. Do not pre-digest with summarize; read the actual
   sequence of entries yourself.
4. Only then, if something looks off, use compare_periods / anomalies /
   correlate to verify that specific suspicion. These tools verify a hypothesis;
   they don't discover one.

WHAT MATTERS MOST:
[Name the metrics you trust most — e.g. "My Oura Readiness Score and Sleep Score
are high quality; always check them."] [If you keep a free-text journal metric,
say so — e.g. "I keep a daily journal with context about my day. Read the actual
text, not just whether an entry exists."]

WRITE (under 500 words, prose, no headers or bullet lists):
- Lead with what actually mattered — usually a hunch that moved, a journal entry
  that explains a biometric shift, or an event that gives the numbers context.
- Never report a biometric shift without checking what I logged or wrote that
  week that might explain it.
- Name what I didn't log: call out any daily-cadence metric with no entry in
  2+ days.
- Ask exactly one specific question about something the data can't show you —
  not "how are you feeling," but something like "you logged pain twice last week
  but not since Tuesday — did that resolve?"
- A boring day, or one where I mostly forgot to log, deserves a short response.
  Don't pad it.

DELIVERY:
Work out today's date yourself. Wearable data usually lags overnight, so the day
in focus is the most recent day with a synced score, not necessarily today.
Write the brief as your final response in this conversation — plain prose, no
subject line, as if writing me a letter. I may reply with follow-ups; treat that
as a continuation of the same conversation and use the Cohearence tools to dig
further into anything I ask about.

Why the prompt is shaped this way

Each of those instructions is load-bearing, and knowing why makes it much easier to write your own.

Hunches first. A hunch is a relationship you've saved to keep an eye on — "drinking wrecks my next-night deep sleep" — baselined when you save it and re-checked as new data lands. That makes your hunches a curated attention layer: the things you went out of your way to say you cared about. Lead with them and the brief has a spine that carries across days. Leave them out and the agent picks a fresh topic every morning, so nothing ever builds.

Don't rank by activity count. Metrics synced from a wearable land every single day and dominate any ordering by recency or volume. The things you log by hand — mood, pain, a note about your evening — arrive a few times a week and sink to the bottom. Say so explicitly, or the interesting half of your data gets buried under sleep scores.

Pull raw, don't summarize. This is the one that most changes how the brief reads. The statistical tools can't see text at all, so a brief built on summarize and compare_periods silently drops every journal entry and note you've written — which is exactly the context that explains the numbers. A month of raw measurements is a small amount of data, and reading the actual sequence of events is what makes a brief feel like it's about your life rather than about a dashboard.

Verify, don't discover. correlate, anomalies, and compare_periods answer "is this true about the metric I named." They're for checking something the agent already noticed by reading. Lead with them and you get a statistics report; save them for confirmation and you get an observation with evidence behind it.

Making it yours

• Personality. The voice above is one option. Blunt, funny, clinical, gentle — it's a line of prompt, so try a few.
• Coverage. Add whatever blocks earn their place: a workout recommendation grounded in your recent readiness and training load, a closing reflection with one concrete thing to change, a weekly variant that reviews hunches over 90 days instead of days.
• Cadence. A short daily brief and a longer Sunday review are different prompts and different schedules. Running both is reasonable.
• Where it lands. By default the brief waits in the run's session. Attach another connector and the agent can email it, post it to Slack, or write it into a document instead.
• Writing back. Ask the agent to propose a hunch when it notices something recurring, and the check-in starts to accumulate instead of resetting each morning. See By MCP for how proposals get approved.

---

Built a check-in that works well? Email us at [email protected] — we'd like to see it. For the other direction, where a scheduled agent writes data into Cohearence, see By Automation.