close-week git:20261030.6b6d663A

Immutable. This exact content never changes and is served at /api/v1/blob/0471664fadc528bb.

---
name: close-week
description: Closes the workweek and sets up the next one. Aggregates the daily balances, reads the week's pattern, forces 1 to 3 decisions, revisits the recorded pains, measures the health of the system's memory, keeps a trajectory signal and leaves the new week's journal ready. Use when the person says "close the week", "close my week", "weekly review", "how was my week", "plan next week", "/close-week", or the 0.7.0 phrases "fecha a semana", "fechar a semana", "review da semana", "como foi minha semana", "monta a próxima semana", "/fechar-semana", usually on the afternoon of the last workday. On Friday at the end of the day, suggest or run the skill even if the person does not ask: the daily closes already ran all week and that is when the week really closes. Do NOT use to close a day (close-day), start the day (start-day) or for a midweek update (journal directly).
---

# Close the week

> pm-chief-of-staff seed. This file is YOURS: ask the agent to change any step and it edits this skill on the spot.

Writing: write to the person and in their files in the language this folder's instruction file defines (the `AGENTS.md`; in a folder without `AGENTS.md`, the `CLAUDE.md`); without it, in the language they write in, and fallback questions go in their language. Follow that file's canon. Without it: no em dash (use a comma, parentheses or a colon), short dense sentences, 1 line per item.

`close-day` assesses a day; this skill reads the whole week and decides. Day to day lets slip what only shows up in series: the same reason 4 times, the same task slipping, the priority stalled for 10 days. Compiling is pulling what the daily closes already recorded; interviewing is the exception.

## Flow

1. **Compile the week from what is written**: the `## Today's 3:` blocks in the journal (the sequence of `**Balance:** N/3` + reasons), the `**Reflection:**` lines, the `memory/` logs of each day, the diff of the folder, and the previous `## Week close:` if there is one (trend: the pattern that repeats weighs more than the one of the isolated week). No balances because the days were not closed: compile from what there is and say that the pattern gets poorer without the daily close.
2. **Priorities**: for each priority in `CRITICAL_FACTS.md`, answer with evidence: moved, partial or stalled? Closed for good: propose the update to the file. Stalled 2+ weeks: goes into the calls. Before writing to a shared destination (here and in the next steps: hub, journal, `CRITICAL_FACTS.md`), read `.claude/skills/_shared/write-gate.md` and apply the admission decision (ADD, UPDATE, SUPERSEDE, NOOP). Without the module, the check in 1 line: read the destination before writing, and do not write what is already there.
3. **Read the pattern in the aggregated reasons**: the same reason on 3+ days is a self-management pattern, and each points at a different adjustment: recurring **out of time** = ambition miscalibrated against the calendar · **outside attack** = a leaky acceptance filter or a structurally reactive week · **wrong 3** = the morning prioritization missing the target. Close with the week in 2-3 sentences: focused or reactive, and did the deliveries serve the priorities?
4. **(If you lead) pass over the team**: who delivered, who slipped repeatedly, and a direct report's task you ended up taking on (add up the hours: it is the week's Super-Leader showing up in numbers).
5. **Pains**: does `context/pains.md` exist? Reread it: did any improve, get worse or pass its review date? An overdue pain with no movement becomes a call.
6. **Memory health**: `wc -c` on `CRITICAL_FACTS.md` against ~3KB (the ~700 tokens of the target, in bytes) and on the instruction file (the one defined at the top of this skill) against ~10KB; these files enter every session, and a yardstick without measurement inflates in silence. The two numbers are curation targets, not technical limits: the session-start hook has its own, much higher ceiling (32KB per file), for another purpose. Within the target, the step stays silent in the answer. Over it: propose concrete curation, never cut alone. In `CRITICAL_FACTS.md`: an event that stopped being active becomes history (log or hub) and a resolved priority leaves. In the instruction file: a rule that belongs to a specific skill migrates to the corresponding planted skill (it is the person's, editable); a rule that looks dead, the person decides whether to kill it. A folder with `AGENTS.md` and `CLAUDE.md`: check that `CLAUDE.md` is only the `@AGENTS.md` line. An extra line there is a rule Claude reads and another agent does not: propose in 1 line moving it to `AGENTS.md`, and never move it alone; a `CLAUDE.md` with only the import stays silent in the answer. In the same pass, measure in the hubs touched during the week the head (Facts + Current reading) against ~half a page (~15 lines), same status of curation target; over it: propose which line becomes history. A hub with "Recent signals" piling up old signals already consolidated in the body calls for pruning: propose it in 1 line. A dated prospective item ("for the next conversation", a follow-up with a date) whose date already passed is never carried forward in silence: propose its expiry, moving it to "History" with the mark "expired on YYYY-MM-DD" (the mark separates what lapsed from what was done); only what has a written date expires, with no inference of lapse. If a module exists in `.claude/skills/_shared/`, check that each skill declared at its top (the `**Loaded by:**` line, or `**Carregado por:**` in an 0.7.0 folder) still references it: a consumer without the reference stops applying the discipline without raising an error, and that top line is the only clue. The skill measures and proposes; the decision is always the person's.
7. **Archive candidates**: list hubs and notes with no write in 4+ weeks in `people/`, `projects/` and `reference/` (cheap detection: modification date via `ls -lt`, or the last dated line of the file). A list with more than ~10: show the 10 most stalled and say how many were left out. Write silence also happens in a stable hub of a person who is still active, so the list is only of candidates and archiving is the person's decision, never automatic. A week with no candidate leaves the step out of the answer.
8. **Trajectory**: 1 to 2 signals of your impact in the week (a delivery that moved a number, a hard decision taken, recognition received) → propose 1 dated line in `context/trajectory.md`. It is the material that the cycle review and the promotion conversation ask for, and that nobody rebuilds from memory later; whoever does not lead is the one who most needs the record.
9. **The calls (1 to 3, mandatory)**: decisions named from what showed up, in the format fact → decision: "{priority} has not moved for 2 weeks: reprioritize it or admit it died" · "the reason {X} closed {N} days: {adjustment from step 3}" · "{task} slipped 2 weeks: do it, delegate it or kill it". A week with no call calls for distrust (a shallow review is more common than a perfect week); if the week really was good, the call is "repeat {what worked}".
10. **Set up next week**: create the new journal in `journal/`, in the naming pattern the folder already uses. Migrate every `- [ ]` still relevant, deduplicated and dated; a task slipping 2+ weeks only migrates with the step 9 decision taken. A prospective item anchored in a date that already passed does not migrate by inertia: propose expiring it (the line stays in the closing journal, marked "expired on YYYY-MM-DD") or re-anchoring it with a new date, and the person settles it; an item with no written date stays outside this rule. Promote the week's last `**→ tomorrow:**` to Monday's first task: it is the link of the turnover, without it Friday's seed is orphaned in the old journal.
11. **Record and answer short**: the analysis (pattern + calls) appended at the end of the closing journal, under `## Week close: YYYY-MM-DD to YYYY-MM-DD`; in the chat, the pattern in 2-3 sentences + the calls + the links to both journals. Do not retell the week.

## The loop

`start-day` and `close-day` turn the day; this skill turns the week: it consumes the daily closes, decides what the day does not decide and leaves the new journal ready for Monday's `start-day`.

## When NOT to use

End of an ordinary day → `close-day`. Start of the day → `start-day`. Middle of the week ("did X") → direct update in the journal.