End the Day

Purpose

A closing-out pass, not a rewrite. The goal is to catch action items that would otherwise fall through the cracks (jotted in a note but never actually tracked anywhere, or left as an unresolved open question) and offer sharper wording where it helps - while leaving every actual edit, and every change to your task tracker, to your approval. Never silently tidy, restructure, or rewrite prose.

end-task is the lighter counterpart - a quick per-chat check that Claude's own memory files reflect what the session just did, cheap enough to run every time a task chat closes. This skill is the heavier once-daily version: a full sweep across every note edited today, plus completed-tasks logging and brag-item suggestions. Run end-task when closing an individual task chat; run this once, at the actual end of the day. (An earlier version folded both into this skill; they were split back out because a full vault sweep is too heavy to run every time a chat closes.)

Where "tracked" comes from

This skill needs to judge whether a bullet is already tracked somewhere, and that depends on how you track tasks (see start-the-day for the fuller discussion of vault-native vs. external task managers):

• Vault-native: if you're using Obsidian's Tasks plugin (or similar), "tracked" means a - [ ] checkbox carrying whatever tag/marker your setup requires (e.g. #task). Judge against what's actually in the note, not an export.
• External task manager: if tasks live in a separate app, pull its current export (see skills/things3-export/ for a Things3 example, or your own equivalent) and judge against that instead - a plain bullet in a note isn't "tracked" just because it reads like an action item.

Whichever applies to your setup, use that as the source of truth for step 2 below; don't invent a third, vault-checkbox-shaped judgment when you're actually using an external tracker, or vice versa.

Out of scope now: a project-drop-folder review and a public-repo sync check both used to live in this skill. The drop-folder review didn't hold up in practice, and the repo sync was slow enough to hold up finishing for the day, so both were pulled out. Bolt either back on as your own pass if the pattern is useful to you.

Trigger

/end-the-day, or asking to wrap up / end / close out your day. Not triggered by closing an individual task chat - that's end-task.

Process

1. Find today's edited notes.
   • Always include today's Daily note.
   • Also check the vault for any other file modified today (mtime), so the scan isn't limited to the daily note if you edited a project note, client note, decision log, etc.
2. Pull whatever represents "tracked" for your setup (see above) - the current task-manager export if you use one, or just the notes themselves if you're going vault-native - so loose-end detection is judged against what's actually tracked, not guessed at.
3. Scan those files for loose ends:
   • Plain bullets that read like a concrete action item but don't obviously correspond to anything tracked - flag them and ask if the user wants each one added. If they say yes to a specific item, add it to the task manager's inbox (for Things3, via AppleScript's make new to do) with a title and optional note only - never a project, area, tag or date, so the inbox stays the one place new tasks get sorted. Never add anything without that per-item yes. If you'd rather this skill stayed strictly read-only, drop the adding and just flag the item.
   • Open questions or unresolved "?"/"TBD"-style notes - flag as "still open?" rather than assuming resolved.
   • Likely duplicates - the same action mentioned in more than one note, or a note-jotted item that's clearly the same thing as an existing tracked task worded differently.
4. Suggest sharper, more action-oriented wording for any flagged item that reads vague or passive (a bare topic instead of a verb-led instruction). Offer the reworded version alongside the original - don't apply it.
5. Report findings, each item tagged with what kind of issue it is (unlinked action item / open question / duplicate / wording suggestion), with a one-line reason. Keep it scannable - this is a closing-out pass, not a report to read carefully.
6. Ask before changing anything, one item at a time. Use structured question/answer prompts (Yes/No, plus a free-text option) rather than a single wall-of-text list - reviewing one decision at a time beats a big block to respond to all at once. Work through findings this way rather than batching them into one message. No silent edits, no silent changes to task state.
7. Run the log-completed-tasks skill for today, writing a "Completed today" section into today's Daily note. This is a factual pull from your task manager, not a judgement call - present its output as-is, and don't treat it as a complete account of the day's work (plenty gets done that never becomes a task). If you track tasks inside the vault rather than a separate task manager, skip this step and step 8.
8. Auto-detect a sparse day, and if so, scan for real work the task manager never saw. Some days are spent mostly in Claude chats or notes and barely touch the task manager, so step 7's list badly undersells them. Check for that: fewer than 3 completions from step 7 AND 2 or more non-Daily-note files edited today (from step 1), with extra weight if any of those are a decision log or project note. If the condition isn't met, skip this step entirely and don't mention it - don't force a scan onto a normal or quiet day.
   • If it is met, scan the step 1 files for entries that read as genuinely finished work - a decision reached, a build step completed, a call that resolved open questions - with no task completion behind it.
   • Draft each as a short, standalone bullet (same style as suggest-brag-items drafts - understandable out of context, linked back to its source note).
   • Present all drafts together as one checkbox list (multiSelect) so the user ticks which ones actually happened. If your question tool caps the number of options, split into several rounds rather than silently dropping the excess.
   • Add only the ticked bullets to the "Completed today" section, clearly separated from the task-manager list (e.g. under a sub-line like "From today's notes and calls (not in the task manager):") so the section's provenance stays honest - never blend an inferred item into the factual list.
   • If nothing looks like a real gap, say so briefly and move on.
9. Run the suggest-brag-items skill over the same file set gathered in step 1, so it isn't re-scanning the vault separately. Present its output as-is - this is a separate concern from loose-end detection (achievements worth keeping, not action items that fell through the cracks), so keep it as its own pass. Some items from step 7 or 8 may also surface here as brag candidates - that's expected overlap.
10. Verify before reporting the day closed. List every file actually created or modified this session, re-read each one, and quote the specific lines that prove the change landed. Flag anything claimed as done earlier in the session but not actually reflected in the file. Don't skip it even on a quiet session with few edits.
11. File the daily chat (optional, Claude desktop app only). The very last thing, after verification. Move this chat into a sidebar group named Dailies using the desktop app's session tools, looking the group up by name each time rather than hardcoding its id. If no such group exists, say so and stop - don't create one. If the tools aren't available (e.g. a terminal session), skip silently. Done here rather than in start-the-day (which names the chat) so today's chat stays in the main list while it's in use, and the Dailies group only holds finished days and can stay collapsed.

Out of scope

• Rewriting prose/notes for clarity beyond flagged action items.
• Reorganising sections or merging notes.
• Judging whether a task is actually done, or marking anything complete in a task tracker - that's the user's action, this skill only surfaces and asks.
• Organising, scheduling or tagging tasks - the only write is adding a new inbox item on a per-item yes.
• Silent edits of any kind - this skill only proposes.
• Project drop-folder review and repo sync - see Purpose above.
• Per-chat memory freshness checks - that's end-task's job.

Notes on conventions (see the vault's CLAUDE.md for the full set)

• If you're using an external task manager, task tracking lives there, not in the vault - judge against its export, not vault checkboxes.
• If you're vault-native, a checkbox only counts as tracked if it carries whatever marker your setup requires (e.g. #task).
• Flat tags only, never nested, for Obsidian notes.
• No em-dashes in anything written into notes - hyphens instead (edit to taste).