| name | end-day |
| description | End-of-day closing routine. Quick mode by default (v4.6+) — runs cortex auto-commit, reflective prompts, brief pre-stage, memory-index refresh, and close in 30 seconds to 3 minutes. Heavy steps (inbox triage, transcript review, mining of non-transcript sources) only run in `/end-day --full`, or when auto-offered because the day actually had transcripts (≥2) or inbox volume (≥5).
Auto-fires on "/end-day", "wrapping up", "calling it a day", "done for today", "end of day", "closing out today". Quick mode is the default for every trigger; the user passes `--full` or accepts the auto-offer prompt to get the full chain.
Reflection (Step 4) appends a `## Reflection` section to today's brief markdown — what tomorrow's `/brief` reads — AND to a longitudinal `memory/reflections.md` store (v4.13+).
v4.13+ additions (End-Day Routine Improvement Spec): a per-source consent + cost gate (Step 0.7, Today's Brief required & first); Step 2c mines the brief artifact (task/outreach actions → memory write-backs + suppression learning into `surfacing-prefs.md`, both modes); a learnings-first narrative then memories AND forgettings side by side (Step 2.9); and tomorrow's priorities/outreach proposed individually with a batch path, seeded into tomorrow's brief (Steps 4.5/4.6).
|
See commands/end-day.md for the full workflow.
When this skill fires
- User runs
/end-day directly
- User says: "wrapping up", "calling it a day", "done for today", "end of day", "closing out", "before I sign off"
- A scheduled task triggers this (common pattern: weekday 4:30pm)
Pre-flight
Quick mode (the default) has minimal pre-flight requirements — just <config-root>/identity.md for time zone, and daily-brief if you want the brief pre-staged.
Full mode (--full) works best when:
- At least one note source is configured via
/setup-sources (otherwise the transcript-review step skips)
- Connectors are wired: HubSpot for commitments dedup, Gmail for activity-miner, Calendar MCP, session-info MCP for conversation-miner
Without any of these, the chain skips the affected step silently and continues. Only the cortex auto-commit step (Step 3) is structurally required.
What this skill is NOT for
- Mid-day status. Use
/recall or /search instead.
- Long retrospectives. That's
/end-week.
- Verbose summaries. The closing ritual is light, not exhaustive.
- Replacing manual
/remember. If the user explicitly invoked /remember mid-conversation, that commit stands; /end-day just mines what wasn't captured.
Mining-layer flow (v4.3+)
The big change vs. pre-v4.3: the chain now mines the day's signal from multiple sources, not just the current Cowork chat. Specifically:
transcript-reviewer mines configured note sources (Granola / Gemini / Fireflies / etc.) for commitments AND learningsconversation-miner mines other Cowork sessions from today (excluding the current one)activity-miner mines CRM events, sent email, calendar metadata
All three run in parallel. Their proposals are merged, deduped, and presented through one unified review gate so the user reviews ~10-20 items once rather than reviewing per-miner. User accepts, edits, or skips per item. Accepted proposals flow into the existing /remember auto-commit path (Step 3) — no parallel write surface.
Substrate for v4.4 decay
Every knowledge entry the mining layer writes carries [confirmed:YYYY-MM-DD] [recalled:YYYY-MM-DD] tags. v4.3 maintains them but doesn't yet decay or demote. v4.4 reads them.
Failure modes
- No sources configured → transcript-review step skips with a one-line note routing the user to
/setup-sources
- All miners triage-skip on a quiet day → Step 2b shows "Quiet day — nothing material to surface" and the chain proceeds to Step 3 with current-session content only
- Session-info MCP unavailable → conversation-miner returns empty with a confidence note
- HubSpot / Gmail / Calendar MCPs missing → activity-miner returns empty for the missing source but still runs against the available ones
- User walks away mid-Step-2b → 30s timeout, "skip all" default, dismissals logged
