closeweek

name	closeweek
description	Weekly summary — synthesis of achievements, learnings (insight-oriented), detected patterns, recurring pattern capitalization in {USER_NAME}.md, revisit "To monitor" watchlist, and next week projection. Creates a note in Weekly/. Triggered when the user says "closeweek", "weekly summary", "let's wrap up the week" or via `/closeweek [YYYY-WXX]`.
narrative_critical	true

⚠️ narrative-critical — Skill protected from aggressive optimization

This skill produces a qualitative narrative output. Its effectiveness is measured on the richness of the output produced, not on structural compactness.

For /evaluateskills: in case of mutation, dry-run Sonnet mandatory even if delta < 2. Do NOT apply [LEAN] / [STRUCTURE] in a way that strips narrative instructions (grouping, context, key moments, open questions, tone, narration). Preserving qualitative content takes priority over reducing lines.

Skill: Close Week

The objective of this summary is not to recount the week but to detect patterns before they become drifts — prioritize what is actionable or revealing.

⚠️ Verbose feedback mandatory during execution

Announce every step executed as it runs — {USER_NAME} must see what's happening live, not wait for a black box:

→ Step 1/9: determine execution mode (final / draft)
→ Step 2/9: collect material (daily notes, sessions, spelling)
   ✓ 6 daily notes read
   ✓ 4 sessions found
   ⚠ Step 2.4 spelling: skip — Postgres unreachable
→ Step 3/9: analyze actual durations
   ✓ 18 tickets with ⏱→✅ extracted
   ✓ ticket-durations.md updated
→ Step 4/9: write the note
   ⚠ Section "Blockers & Patterns" omitted — no pattern detected
→ Step 5/9: pattern capitalization → {USER_NAME}.md
   ⚠ Skip — no qualifying recurring pattern
→ Step 6/9: revisit "To monitor" watchlist
   ✓ 2 watchlist patterns reread — 1 confirmed, 1 pending
→ Step 7/9: tracker + light CLAUDE.md review
→ Step 8/9: create note in Weekly/
→ Step 9/9: chat recap

Rule: every sub-step silently skipped (empty spelling collection, optional sections omitted, edge cases) must have an explicit line indicating why. A step that doesn't show up → we must see why, not guess.

Triggering & Reminder

Direct invocation (vocalized):

• {USER_NAME} says trigger: "closeweek", "weekly summary", "let's wrap up the week", "what's the summary"
• Claude Code detects and launches the skill automatically

Manual invocation (command):

• /closeweek alone → processes current ISO week (calendar week of today)
• /closeweek 2026-W14 → processes specific week (ex: previous week for catch-up)

Automatic reminder:

• During /today: check command-tracker.md → if gap > 7 days since last /closeweek, display message "Last closeweek: W##, N days ago. Restart?" (don't automatically trigger closeweek, just notify {USER_NAME})

---

Step 1 — Determine execution mode

Calculate target week: argument YYYY-WXX passed or current ISO week by default. Deduce Monday (D0) and Sunday (D6) to delimit files to consult.

Decide flow based on situation:

Condition	Action
Argument earlier than current week (ex: now W14, argument W12)	Final mode: create/overwrite YYYY-WXX.md
Current week (today ∈ W) AND today = Sunday	Final mode: create/overwrite YYYY-WXX.md
Current week (today ∈ W) AND today ≠ Sunday	Draft mode: create YYYY-WXX-draft.md with header "(Partial summary — finalize on Sunday)". Continue normally after creation.
Note YYYY-WXX.md already exists (final mode)	Display existing content. Propose to {USER_NAME}: [A] add missing elements, [R] replace completely, [C] create new version (YYYY-WXX-v2.md). Wait for response.
Missing weeks detected (between last closeweek and target)	Display list in ascending order. Propose: [T] process missing weeks first or [S] skip and process target week. Wait for response.

In case of missing weeks validated [T]: process the oldest first (call closeweek recursively on each), then return to target week.

---

Step 2 — Collect material

Sequential reader (in this order):

1. Read last closeweek prior to target in {VAULT_PATH}\{DAILY_NOTES_FOLDER}\Weekly\ → extract unclosed threads and recurring learnings
2. Read daily notes of the week in {VAULT_PATH}\{DAILY_NOTES_FOLDER}\ (files YYYY-MM-DD.md in target Monday-Sunday range)
   • Extract from each note's frontmatter: energy, score, feeling (if present) → store for Energy section (Step 4)
3. Read sessions of the week in {VAULT_PATH}\{CLAUDE_CODE_FOLDER}\Sessions\ (files YYYY-MM-DD.md in target Monday-Sunday range)

Edge case: 0 daily notes found

• Display: "0 daily notes found for W##. Proceed?"
• Propose: [P] proceed from sessions alone or [A] cancel summary. Wait for response.
• If [P]: display "Lightweight summary (sources: sessions only)". Continue with Step 3 (duration analysis) then Step 4 (writing).

Classify material by observable impact:

Level	Definition	Action
1	Structural decision (changes orientation/org/direction)	Always include
2	New pattern/workflow/habit detected	Always include
3	One-off achievement (contextual insight)	Include if relevant
4	Routine (no notable surprise)	Omittable

Unclosed threads and recurring learnings surface at the top of the summary.

Edge case: Empty week (no level 1-2 content)

• Write short minimalist summary (3-4 items). Include line: "Stable week — few events or consolidated routines."

If many sessions (3+ per day): synthesize by day or by project (ex: "Monday: Project A setup + tests", "Wednesday: Project B + meeting"). Don't enumerate every session.

Spelling collection

In parallel with main material reading — result stored for Step 9.

1. Sessions Postgres — fetch sessions of the week:

psql postgresql://claude:claude@docker-host:5432/claude_sessions \
  -t -A -F'|||' \
  -c "SELECT date, title, raw_md FROM sessions WHERE date BETWEEN '[Monday YYYY-MM-DD]' AND '[Sunday YYYY-MM-DD]' ORDER BY date, time"

If 0 results → ignore, continue.

2. Daily notes reflections — extract ## 💡 Ideas & Reflections sections from daily notes already read this week (exact title from {USER_NAME} template). Concatenate into single text block.

3. Parallel Haiku agents — launch 1 agent per session raw_md + 1 agent for aggregated Reflections block. Exact prompt:

"Identify spelling mistakes, incorrect agreements or faulty turns of phrase in this French text. Return only a JSON list: [{"erreur": "...", "correction": "...", "regle": "..."}]. Maximum 5 items. If no mistakes detected → return []."

4. Aggregate — merge all JSON outputs, count occurrences of each mistake, retain 2-3 most frequent → variable ortho_items. If ortho_items empty → silently omit in Step 9.

Long-term recurrence detection (optional, silent fail)

For each level 1-2 pattern identified in material (Detected patterns, Blockers):

uv run ~/.claude/semantic_search.py "<pattern in 3-5 words>" --top-k 3 --since <Monday-4weeks>

If similarity ≥ 0.45 on sessions prior to target week → cross-week recurring pattern:

1. Age: date of first occurrence found
2. What resulted: rerun --type completed --since <first occurrence date> to see follow-ups
3. Store in variable pattern_recurrents: list of {pattern, first_date, follow_up}

If Postgres/Ollama unreachable → pattern_recurrents = [], non-blocking, display: ⚠️ Recurrence detection unavailable (Postgres/Ollama down) — patterns sourced from prior closeweek only.

---

Step 3 — Analyze actual durations

Why: since the /workon patch of 2026-05-03, each ticket carries ⏱ HH:mm→HH:mm ✅ in the daily note. These observed durations calibrate future /today estimates (which reads ticket-durations.md at every plan generation). Without this step, /today stays with rough estimates (45min default).

Source: the daily notes of the week already read in Step 2.2. Don't reload — reuse content in memory.

3.1 — Extract durations

For each daily note of the week, parse checked lines with start + end timestamps:

• Pattern: - [x] ... ⏱ (\d{2}:\d{2})→(\d{2}:\d{2}) ✅
• Cross-midnight: if end < start (ex: ⏱ 23:45→00:30) → add 24h to end before subtraction. Duration = 45min, not -23h15.
• Duration: (end - start) in minutes.

Ignore lines without the arrow → (old formats ⏱ HH:mm alone or ✅ HH:mm separate) — no extractable duration, don't try to infer.

3.2 — Classify by type

For each extracted ticket, determine its type from the [[TicketName]] link and line context:

Type	Detection signals
CC maintenance	Ticket in 99 - Claude Code/Claude Code Kanban.md — patch hook, config, lessons, cleanup
CC skill	Ticket in 99 - Claude Code/Claude Code Kanban.md — skill creation or overhaul
FSTG	Ticket in 04 - Projects/From Sprue to Glory/
HUSKER	Ticket in 04 - Projects/Husker/ or slug starts with husker-
GYLT	Ticket in 04 - Projects/Get Your Life Together/ or slug starts with gylt-
Perso practical	Section ### 🎯 Perso outside dev projects (admin, courses, tidying)
Hobby	Section ### 🎨 Hobby
Vault maintenance	Section ### 🔄 Vault maintenance (vault commands /harvest, /link, etc.)

If a ticket doesn't clearly match any category → type Other. Don't invent sub-types.

3.3 — Calculate stats by type

For each type represented this week:

• Nb tickets: number of entries
• Average: sum of durations / nb tickets
• Median: ascending sort + center value (or average of 2 centers if nb even)
• Min / Max: extremes observed this week

Format durations as Xmin if < 60min, else XhYmin (ex: 1h20, 45min, 2h).

3.4 — Update ticket-durations.md

Path: {VAULT_PATH}/{CLAUDE_CODE_FOLDER}/ticket-durations.md

If file doesn't exist → create it with header + initial table.

If file exists → don't overwrite. Merge:

1. Read existing table (historical types)
2. For each type present this week: recalculate cumulative stats (history + new tickets) — ticket-durations.md keeps individual tickets at end of file, reread them to recalculate.
3. For types untouched this week: keep as-is
4. Append section "Individual tickets (week WXX)" at end of file (raw history, never purged)

Format:

# Real ticket durations

*Updated on YYYY-MM-DD (closeweek W##)*

Data calibrated from tickets `⏱→✅` in daily notes since 2026-05-03.

| Type | Nb tickets | Average | Median | Min | Max |
|------|-----------|---------|---------|-----|-----|
| CC maintenance | 12 | 28min | 25min | 10min | 45min |
| CC skill | 5 | 52min | 45min | 30min | 1h20 |
| Perso practical | 4 | 1h45 | 1h45 | 1h30 | 2h |
| ...

## Individual tickets

### Week W18 (2026-04-27 → 2026-05-03)
- 2026-04-28 — [[ticket-x]] — 32min — CC maintenance
- ...

### Week W19 (2026-05-04 → 2026-05-10)
- ...

3.5 — Edge cases

• No ticket with ⏱→✅ this week: don't touch ticket-durations.md. Log in verbose feedback: ⚠ No timestamped ticket this week — skip ticket-durations.md update
• Type with < 3 tickets in total: keep line in table but it's /today that will detect insufficiency (signal "insufficient data for [type]" on /today side, not here)
• File ticket-durations.md corrupted (parsing fail): don't overwrite. Log ⚠ ticket-durations.md unreadable — skip update. Manually rebuild. Continue with rest of closeweek.

---

Step 4 — Write the note

Template:

# Week WXX — YYYY
*From [Monday DD/MM] to [Sunday DD/MM]*
*[Partial summary — finalize on Sunday]* ← add if draft mode

## Continuity from previous week
[Unclosed threads, recurring learnings from W-1]
← Omit this section if no thread to pursue

## ✅ Achieved
- ...

## 💡 Learnings
*Formulate as "X reveals that...", "X means that..." — not bare factual recap.*
- ...

## 🔁 Blockers & Patterns
*What worked, what blocked, emerging patterns.*
- Worked: ...
- Blocker: ...
- Pattern detected: ...
  *If recurring pattern (variable `pattern_recurrents`):*
  `→ Recurring since YYYY-MM-DD — follow-up observed: [what resulted, 1 sentence]`
← Omit entirely if no blocker/pattern this week

## 🔗 Connector theme
*Meta-pattern linking achievements/blockers in 1 sentence.*
← Omit if absent

## 🔋 Week energy
| Day | Energy | Score | Feeling |
|------|---------|-------|---------|
| Mon | [X]/5 | [X]/5 | *[feeling]* |
| Tue | ... | ... | ... |
| Wed | ... | ... | ... |
| Thu | ... | ... | ... |
| Fri | ... | ... | ... |
| Sat | ... | ... | ... |
| Sun | ... | ... | ... |
← Omit days with no daily note. Omit Score or Feeling column if no daily note this week contains that field.

## ➡️ Next week
- ...
← Omit if empty

← Omit entirely if level 0

Writing rules:

• Optional sections (Continuity, Blockers & Patterns, Connector theme, Next week): omit silently if empty — never leave section title empty
• Insight > recap: each point must imply something for what's next
• Little content (< 3 items total): short summary + explain why (ex: "Stable week — few events")

---

Step 5 — Pattern capitalization → {USER_NAME}.md

Objective: anchor recurring patterns detected this week in 01 - Me/{USER_NAME}.md §Observed patterns — without waiting for the monthly.

Eligibility criterion: a pattern is proposed for anchoring if at least one of these conditions is true:

• Present in pattern_recurrents (variable Step 2 — semantic recurrence across weeks with similarity ≥ 0.45)
• Mentioned in ≥ 2 distinct daily notes of the week (similar phrasings even if not identical)
• Linked to an unclosed thread resumed from W-1 or W-2 (continuity observed over 3 weeks)

If no qualifying pattern → skip silently, verbose line: ⚠ Skip — no qualifying recurring pattern. Go to Step 6.

If qualifying patterns:

1. Read {VAULT_PATH}\{PERSONAL_FOLDER}\{USER_NAME}.md — identify the most recent ### Observed patterns section (by date).
2. Present to {USER_NAME}:

Recurring patterns qualified this week:

1. [Pattern short — 1 sentence]
   Occurrences: [daily YYYY-MM-DD, daily YYYY-MM-DD, pattern_recurrents since YYYY-MM-DD]
   Proposal for anchoring in {USER_NAME}.md §Observed patterns (current month):
   → "[Proposed formulation, 2-3 sentences with dated sources]"

2. [Pattern 2 — ...]

3. Wait for validation {USER_NAME} for each pattern. Accept: "1", "all", "1 and 3", "yes", "ok". Refuse: "no", "skip", silence/timeout (45s) → skip this pattern.
4. For each validated pattern:
   • If section ### Observed patterns ([current month]) exists → add bullet under that section
   • Otherwise → create section ### Observed patterns (closeweek W## — YYYY-MM) at head of Patterns zone (before existing monthly sections)
   • Bullet format: - **[Pattern short]** — [Description, 2-3 sentences with dated sources]. Pattern to monitor: [optional — falsification condition or next verification]
5. Always pair-programming: no writing without explicit validation. Global refusal ("no", "skip all") → write nothing, verbose line ⚠ Proposed patterns refused — no {USER_NAME}.md anchoring.

If {USER_NAME}.md missing/unreadable → display ⚠ {USER_NAME}.md inaccessible — pattern capitalization deferred, don't block rest of skill.

---

Step 6 — Revisit "To monitor" watchlist

Objective: revisit patterns marked To monitor in {USER_NAME}.md — confirm, refute, or leave pending. Prevents watchlist from bloating without ever being resolved.

Source: {VAULT_PATH}\{PERSONAL_FOLDER}\{USER_NAME}.md

Detection: grep bullets containing one of these phrasings:

• Pattern to monitor
• To monitor:
• to monitor —
• To monitor post-
• To monitor on return

(Case-insensitive. Capture variants while staying strict on the word "monitor".)

If watchlist empty → skip silently, verbose line: ⚠ Skip — watchlist empty. Go to Step 7.

If watchlist non-empty:

1. For each watchlist pattern, present to {USER_NAME}:

Watchlist patterns "To monitor" — reread W## :

1. [Pattern short extracted from bullet]
   Source: {USER_NAME}.md §[section] (anchored YYYY-MM-DD or estimated from bullet context)
   Question: confirmed / refuted / still pending / already resolved elsewhere?

2. [...]

2. Wait for answer per pattern. Accept: "1 confirmed", "2 pending", "1 refuted, 2 resolved", "all pending". Refuse: silence/timeout (60s) → treat all non-responses as "pending".

3. Based on answer:

   • Confirmed → propose bullet modification: remove "To monitor" mention and add Confirmed YYYY-MM-DD (closeweek W##). Wait for OK before edit.
   • Refuted → propose 2 options: [R] remove bullet entirely or [A] archive pattern with note "Refuted YYYY-MM-DD (closeweek W##) — [short reason]". Wait for choice before edit.
   • Pending → do nothing. Bullet stays as-is, will be reproposed next closeweek.
   • Already resolved elsewhere → ask where (other essay, ADR, [[wiki link]]), propose removing "To monitor" mention + add Resolved in [[link]] YYYY-MM-DD. Wait for OK.

4. Strict pair-programming: each {USER_NAME}.md edit awaits explicit validation. Refusal → keep as-is, note in verbose ⚠ Pattern N: edit refused.

If {USER_NAME}.md missing/unreadable → display ⚠ {USER_NAME}.md inaccessible — watchlist not revisited, don't block.

---

Step 7 — Tracker + light CLAUDE.md review

Action 1: Update command-tracker.md BEFORE note creation

Read {VAULT_PATH}\{CLAUDE_CODE_FOLDER}\command-tracker.md. Add or update line:

/closeweek [YYYY-WXX] → executed [YYYY-MM-DD HH:MM]

Action 2: Light review of global CLAUDE.md (advisory, weekly)

Quickly reread ~/.claude/CLAUDE.md (= 99 - Claude Code/CLAUDE.md) + 99 - Claude Code/vault-context.md. Spot any rule obsolete, redundant or misplaced per routing rule: stays in CLAUDE.md = always active all contexts; else → hook (event), skill (invocation) or project context.

• Nothing to report → verbose line ✓ CLAUDE.md reread — no issues.
• Drift spotted → flag it to {USER_NAME} in Step 9 (recap), modify nothing (pair-programming). For deep pass, point to ticket passe-claude-md-inspiration-karpathy-maintenance-reguliere.
• Deep pass cadence: quarterly or after major model update — tracked in command-tracker.md (line claude-md-pass).

Action 3: lessons.md distillation nudge (advisory, weekly)

lessons.md is injected at every prompt by the lessons-loader hook; it's a gateway, not an archive. Count lessons: grep -c '^- \*\*' "99 - Claude Code/lessons.md". If > 20 lessons (re-bloat threshold) → flag to {USER_NAME} in Step 9 (recap): "lessons.md = N lessons, above threshold. Run /lessons-distill to route stabilized lessons to their home (SKILL.md / CLAUDE.md / ADR / Knowledge / prune)." Don't launch distillation during closeweek (heavy pass) — just nudge. Below threshold → verbose line ✓ lessons.md below threshold (N lessons) — no nudge. See [[passe-distillation-lessons-md-graduer-foyer]].

Implicit confirmation: if read + write succeed, continue to Step 8.

---

Step 8 — Create the note

Final path: {VAULT_PATH}\{DAILY_NOTES_FOLDER}\Weekly\YYYY-WXX.md or YYYY-WXX-draft.md per mode (Step 1)

Generate the note with written content (Step 4).

Success signal to {USER_NAME}:

• Display direct link to created note
• Message: "Week W## summary created" (or "draft created — finalize on Sunday")

---

Step 9 — Chat recap

After note creation, display directly in chat (not in note):

1. Narrative recap of the week — 3-5 lines max, oriented on meaning and movement, not task list. Include connector theme if present.

2. W+1 priority table — format:

Priority	Subject
🔴 Urgent/blocking	...
🟠 Important	...
🟡 To advance if possible	...

Source priorities from ➡️ Next week section of note + unclosed threads detected in Step 2.

4. Spelling (if ortho_items non-empty):

### ✍️ Spelling — this week
- "observed mistake" → correction (rule in 1 sentence)

Maximum 3 items. Neutral, factual tone. Silently omit if ortho_items empty.

Rule: the chat recap is always displayed, even in draft mode.

---

Draft → final conversion

Sunday after 23:59 (or {USER_NAME} review earlier):

Run /closeweek 2026-W14 on same week. System detects existing draft (Step 1, condition 4). {USER_NAME} chooses [R] replace completely → new draft replaced by fresh write.

Alternatively, {USER_NAME} can rerun Wednesday and choose [A] add missing to enrich draft then rerun Sunday [R].

---

Absolute rules

• {USER_NAME}.md capitalization: this skill proposes anchoring recurring patterns (Step 5) and watchlist review (Step 6). Monthly (/closemonth) remains responsible for deep consolidation (patterns ≥ 3 occurrences, section overhaul). Weekly capitalizes as-you-go, monthly synthesizes.
• No writing to 01 - Me/{USER_NAME}.md without explicit {USER_NAME} validation — always present proposal before edit
• Process one week at a time if multiple weeks in arrears — propose order (Step 1, condition 5)
• Insight > recap: value = what the week implies, not what it contains

---

Edge cases — Resolution

Scenario	Response
Launched Monday–Saturday current week	Draft mode (Step 1). Note YYYY-WXX-draft.md created. Signal: "Finalize on Sunday: /closeweek or /closeweek YYYY-WXX → [R]".
0 daily notes week	Step 2: display choice [P]/[A]. If [P]: message "Lightweight summary (sessions only)". Continue.
ISO calendar spanning 2 months (ex W13 = Mar-Apr)	Use ISO W13, mention date range in header (ex "From 24/03 to 30/03").
Draft exists for target week	Step 1, condition 4: propose [A]/[R]/[C]. Wait for response before continuing.
command-tracker.md missing/broken	Infer from file creation dates: list all YYYY-WXX.md in Weekly/, extract MAX(mod date), deduce current W. Display "Tracker missing — rebuild from files. Continue?" Wait for confirmation.
ISO calc error (invalid dates)	Display: "Cannot calculate ISO week for [argument]. Verify format YYYY-WXX or date argument. Rerun?" Wait for new argument.
No level 1-2 content (empty week)	Write short minimalist summary (3-4 items) + line "Stable week — few events or consolidated routines."

---

Pair-programming

• No autonomous action: every decision (final/draft mode, missing weeks processing, 0 daily notes procedure) waits for explicit {USER_NAME} validation
• Tracker before note: command-tracker.md update happens first (Step 7), before note creation (Step 8)
• Proposed choices are clear and contextualized: never "Continue?" without explicit option