morning-digest

Use for the daily ~7AM prep — what needs Bryan, meeting prep, today's focus, yesterday's output, and token spend. Fires from a board schedule; also user-invocable.

SKILL.md

GitHub

Morning Digest

Bryan's one morning artifact. Set by him 2026-09-18: "every morning at 7AM you prep me for the day." It replaces the old 05:27 daily-review automated morning run — two briefings ninety minutes apart is the failure this skill exists to prevent, not a feature.

daily-review remains the skill for an intra-day "where are we" pass and keeps its own triggers. This one owns the morning.

How it fires

A board schedule on the morning-digest row of the Team Lead board, not CronCreate. Bryan authorised the switch on 2026-09-18. The rule is a daily calendar rule at 06:47 PT — early enough that the digest is written and pushed by 7, because the gathering takes minutes and a job that starts at 7 delivers late.

Never re-arm this as a session cron. CronCreate is in-memory: it dies on every respawn and auto-expires after 7 days. That is not a theoretical failure — on 2026-09-01 CronList returned "No scheduled jobs" with the re-arm directive sitting unread in context and the last quota reading 50 hours old. The hole it leaves is a missing line, and nothing alerts on an absence. A schedule on the row survives the respawn that kills a cron.

Output shape

The output is today's hit list, written into this week's plan doc. There is no separate digest doc. Changed by Bryan on 2026-10-05: "Instead of the daily digests from Team Lead, every morning can you update this doc and give me a hit list of things I need to do today? Ready by 7AM each morning." "This doc" is the weekly plan, .claude/reviews/weekly-YYYY-MM-DD.md, which is bound to the Team Lead board.

  • Replace the plan's ## Today section each morning. Anchor on the heading, insert the new list, then delete the old items. Never use set_doc_content, because he edits this doc live.
  • The heading carries the date and the total, e.g. ## Today, Tue 10/6 (~45 min).
  • The hit list is at most 7 items, in goal priority order. Each item is something he does today: a decision, a send, a call, a read. Put the time it takes and a link on the noun.
  • Below the list, two optional lines, and only when they apply:
    • Meetings: accepted meetings today, one line each, with the one thing to walk in knowing.
    • Fleet: a token verdict, only when it is not "all clear".
  • The sections below are inputs to the list, not sections of the output. Run the sweeps as written. An item reaches the list only if it needs him today; everything else is dropped.
  • Ranked asks come from the cross-board queue that Team Lead curates. Take the top open asks for this week's goals.

The plan doc lives inside a public repo, gitignored. That is weaker privacy than the old digest path outside the repo. So an item from email or messages names the action and links the thread, and never quotes the content or names a third party beyond what the action needs. Anything more sensitive stays behind the link.

  • Surface the WORKSPACE url, never the bare /review/<docId>. The id lives in .claude/skills/weekly-plan/parent.txt.
  • One push, under 200 characters: the 2–3 things that actually need him today. Not a summary of the hit list — the list is the summary.
  • Every field is its own bullet with a bold label. Prose-with-inline-bold collapses to a wall of text on a phone.
  • Links are Tailscale or GitHub, never *.local, and go on the noun at first mention.

Every item links its own artifact, in every section. Asked for on the first run: "Provide links to Deepcell, learnings-gate-hazards branch, and other unclear items." An item that names a thing he would have to go find — an email thread, a branch, a PR, a board row, a settings pane — carries the link on the noun at first mention.

  • A Gmail thread is https://mail.google.com/mail/u/0/#inbox/<threadId>; the connector returns threadId on every search hit, so there is no excuse for naming a sender without one.
  • Where no URL exists, say so and say what to open. An unpushed branch has no GitHub page — write the branch name and the repo, not a link that 404s.
  • Verify a link resolves before it ships. A dead link costs him more than a missing one, because he only finds out after the context switch.

A setup step that only he can perform ships with its actual steps, not its name. "Include specific step-by-step instructions for Full Disk Access (Messages) and Signal device linking in daily updates." Naming the task is what the digest had been doing, and it leaves him to reconstruct the procedure every morning it reappears.

  • Write the pane path and the clicks, e.g. System Settings → Privacy & Security → Full Disk Access → + → add the terminal app that runs this session.
  • Say where it can be done from. "instructions can be executed from Mac mini, SSH from phone/iPad, or remote desktop." That is per-step and it matters: anything in System Settings needs the GUI, so it is the Mac mini directly or remote desktop — not SSH. A CLI step is fine over SSH from the phone or iPad.
  • Confirm the client before writing Signal's steps. Linking differs between Signal Desktop (scan the QR it displays) and signal-cli (link prints a sgnl:// URI you have to render as a QR yourself). Do not write one procedure as if it covered both.

1. Needs you

Read the last 48 hours. Two buckets, and they are different questions — keep them apart.

  • Unreplied to Bryan — a thread whose last message is not from him. That is the test; do not use read/unread, which tracks whether a tab was open. Say who is waiting and for how long.
  • Worth knowing — anything that changes his day even though nobody is waiting: a cancellation, a deadline moved, money, a reply on something he sent.
  • Say what it wants, not what it says. "Needs a yes/no on Thursday" beats a summary of the paragraph that asked.
  • Three to six items. If the 48h window is quiet, say so in one line rather than padding it with newsletters.
  • Never draft or send a reply from this skill. The digest surfaces; sending is a separate act with its own approval.

The bar is "needs him", not "involves him" (set 2026-09-18)

Three corrections on the first real run, all the same shape: the digest surfaced things nobody was asking him for.

  • A PR is his to review only when HIS OWN LOGIN is on it. gh search prs --review-requested=<user> expands team membership and returns every PR requesting any team he belongs to. Five came back that way on the first run and not one was his. "I think none of these are assigned to me -- don't highlight these in the future unless assigned to me to review." The correct test is the PR's own reviewRequests array — gh pr view <n> --json reviewRequests — checked for his login. Never use --review-requested to decide this.
  • "Worth knowing" is not a licence to include whatever you found. A payment that cleared, a booking already confirmed, a thread someone else closed: if nothing turns on him, it does not go in. "No need to bring these to my attention -- does not need me."
  • Check whether he already handled it before you file it. "Already took care of this." A thread whose last message is not from him can still be finished — by a phone call, a text, a form he submitted. Where the channel cannot show you, mark the item unconfirmed rather than asserting it is open.

Recruiter email goes to the Job Search peer, not to him (set 2026-10-01)

"When you get recruiter emails, pass them to Job search for review. Only highlight for me if it looks interesting. Or they might be well connected and I just want to be on their radar."

  • Send every recruiter thread to the Job Search peer: sender, company, role, how many follow-ups so far, and the Gmail link. One message per digest run with all of them in it, not one message per email.
  • Job Search judges it, not the digest. The digest does not decide that a role is uninteresting; it hands the thread over and stops.
  • It reaches "Needs you" only on Job Search's say-so, for one of his two reasons: the role looks interesting, or the recruiter is well connected and worth staying on the radar of. Name which reason in the line.
  • Read Job Search's board for its verdicts on earlier handoffs before writing the section. A recruiter it flagged yesterday belongs in today's digest even though the email is older than the run.
  • A recruiter's persistence is not a reason to surface it. A third follow-up is noise until Job Search says otherwise.

The Gmail connector is read-and-compose only

Measured 2026-09-18. Search, read and create_draft work. unlabel_thread, label edits, archive and trash all fail with Insufficient scope: required … https://www.googleapis.com/auth/gmail.modify. So this skill can surface a thread and it can never mark one read or clear it — say that plainly when asked, and do not route around it.

update_draft DETACHES a draft from its thread. It returns a new threadId equal to its own messageId, and the reply then arrives as a fresh conversation. To change a threaded draft, create a replacement with replyToMessageId and confirm the returned threadId still matches the original. The orphan cannot be trashed from here — it is Bryan's to delete, so tell him it exists.

A quiet inbox and a failed search read the same. If the search returns nothing, say which query ran and that it ran — "no unreplied threads in 48h (searched newer_than:2d across N threads)". An empty section with no denominator is indistinguishable from a broken one.

Which channels this sweeps

Answered 2026-09-18: "Slack, Messages and Signal. Research if there's a clean way to handle WhatsApp." All four are wanted. Three have a path; one does not yet.

Which channels are in was his call, not a technical one. Do not widen the sweep on your own judgement, and do not narrow it either.

  • Gmail — on now.
  • Slack — on now. The connector was already authorised, so this needed nothing from him. Same unreplied test as email: the last message in the thread or DM is not from him. Sweep DMs and channels where he is @-mentioned — not every channel he belongs to, which is a newsletter, not an inbox.
  • Messages (iMessage) — approved, blocked on a Full Disk Access grant only he can make. No connector exists; the history is a local SQLite store under his Library. That grant is machine-wide rather than scoped to Messages, so never perform it, script it, or prompt for it mid-digest — the steps live on the channel-scope row and he does it in System Settings.
  • Signal — approved, blocked on him linking this machine as a second device. There is no API, by design; signal-cli works by linking, which means a QR scan from his phone and real message history stored on this machine.
  • WhatsApp — not approved. Open research question. He asked whether a clean path exists, which is not the same as approving one. Until one is found and he says yes, do not read WhatsApp by any route. The Business API covers business numbers only; the desktop client's local store and unofficial bridges both breach the terms, with a documented consequence of the number being banned.

Name the channels this run actually read, every time. "Nothing needs you" means nothing in the channels that were swept, and a reader cannot tell which those were unless the line says so. An approved-but-not-yet-wired channel reads as swept otherwise, which is the failure this line exists to prevent.

2. Meetings today

Pull all three of his calendars — the ids are in registry.yaml under calendars:, which is gitignored because this repo is public and a calendar id is a real address. Run list_calendars first and look for a new personal or household one; the other ~16 entries are subscriptions and are not his commitments.

  • Exclude recurring blocks with his partner. His words. A one-off with her is a real meeting and stays.
  • Only meetings he has ACCEPTED — read his own responseStatus. accepted counts, needsAction and declined do not. Organizer counts as accepted.
  • A medication or self-care reminder is not a meeting.

Prep comes from the peer that holds the context, not from you.

  • A person he knows → the Personal CRM peer. Its board id is its workspace_id in registry.yaml.
  • Anything hiring, recruiting or interview-shaped → the Job Search peer, same lookup.
  • Read their boards before messaging them. list_tasks / get_doc answers most of it and wakes nobody. A peer message costs that peer its whole context on the turn it reads, and this runs every day.
  • Message a peer only when a specific fact is missing after reading, and send the question, not the framing. At 06:47 the peer is probably down; spawn it only if the meeting genuinely needs it, and spin it back down after.

Three lines of prep per meeting, maximum: who they are and the last contact, what this meeting is for, and the one thing he should walk in knowing. If you have nothing, write that — invented prep is worse than none.

3. Focus today

At most 5 bullets, in weekly-plan goal priority order. Not by size, not by what is quickest, not by the order you found them.

  • Draw from this week's committed goals, whatever is newly unblocked, and anything now waiting on him.
  • Take it from the weekly plan and the boards. Not Asana — he stopped that on 2026-09-16: "Stop the Asana syncs. It's duplicating stuff that's on the workspaces." Do not read, create, re-date, complete or reword an Asana task from this skill.
  • If your order and the plan's numbering disagree, say so rather than silently picking one. That disagreement means the numbering is stale.
  • Flag an off-track week here, and in the push, by Wednesday. A goal that has not moved at all by midweek is off track regardless of its due date. Name what to drop. Friday is too late — by then the only move left is writing the week off.

4. Yesterday

Exactly 5 bullets on what was actually done. Evidence, not intentions.

  • Merged PRs, closed rows with proofs, docs delivered, decisions he answered.
  • A row closed with no evidence in its comments does not count as done. Say what it was, not that it closed.
  • If fewer than 5 real things happened, write fewer and say the day was light. Padding this section to the count is how a slipping week reads as healthy.
  • Read it from the boards, the transcripts and gh pr list, never from a peer's summary line — a summary is what it meant to do.

5. Token spend

A table of weekly usage against the limit, then one verdict line. Changed by him on 2026-09-18, on the first run: "Show weekly usage vs. limit, not just current pool state" and "Present token spend as a table showing weekly usage vs. limit, not just pool state." Pool state answers "where is the fleet billing right now", which is not the question he is asking.

  • The table is the estate, one row per pool: used against the weekly limit, the Fable meter separately, the reset, and whether the reading is live or stale.
  • Then the runway, measured off points GAINED since the stint began ÷ hours. Never used% ÷ hours, and never projected multiple days off the current 5h window.

The verdict line ends in a decision word: speed up, slow down, or all clear. His words: "Summarize token spend decision rule — speaker needs to know if action is required." A percentage is not an instruction, and he should not have to derive one from the table he was just handed. If the answer is that nothing needs doing, write all clear — that is a verdict, and an absent one reads as an unfinished section.

This digest does not drive a pane for /usage. The live read belongs to /token-watch, on whatever schedule its board row carries. As armed on 2026-09-19 that row fires at 06:00, 47 minutes BEFORE this digest — which makes the panel fresher than it used to be, and does not make it live. Do not restate the cadence as an interval: the armed gaps are not uniform, and deriving one is what made the board's own stall detector file a false alarm on 2026-09-21. Run scripts/fleet_budget_watch.py and report the estate panel as it stands.

So the reading here is stale by construction, and must be labelled. Say "82% or worse", never "82%". A stale figure is a bound, not a measurement — and it can be wrong in either direction, because an idle pool's meter does not only rise. That belief cost two hours at a wall on 2026-09-17.

What not to do

  • Don't send this without the doc bound. A digest that exists only in the terminal is gone by the time he reads his phone.
  • Don't write "blocked on" or "waiting on" from a tmux pane. The pane is a render. Read the session's transcript and find the last turn it actually processed.
  • Don't spawn the fleet to build this. Read boards and transcripts. Waking six peers at 06:47 so they can each report costs more than the digest is worth.
  • Don't let a section fail silently. If Gmail is unauthorized, a calendar did not resolve, or fleet_budget_watch.py exited non-zero, that section says could not run and why. It never says "nothing found". Passed, failed, and could-not-run are three states, and the third one collapsing into the first is the bug.
  • Don't ask him a question in the doc body. Prose is invisible to every queue. If you need an answer, file a review item.