closeday

name	closeday
description	End-of-day wrap-up — synthesize sessions, update daily note with observed patterns, detect WIP hobby in progress, verify essay-check overdue. Trigger when the user says one of these phrases: "closeday", "I'm stopping", "I'm done", "that's finished", "end of session", "good night", "good day", "stopping for today", "that's enough", "day finished", "I'm stopping now", "it's done" or via /closeday.
narrative_critical	true

⚠️ narrative-critical — Skill protected against 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, highlights, open questions, tone, narration). Preserving qualitative content takes precedence over reducing lines.

Skill : Close Day

Trigger

the user says one of these phrases (or equivalent) :

• "closeday"
• "I'm stopping", "I'm done", "that's finished"
• "good night", "good day"
• "end of session", "I'm stopping"

Or: command /closeday

Step 1 — Collect day's data

Target date : calculate before any reading. If hour < 04:00 → date = yesterday, else date = today. Use this date for all paths below.

Read in parallel (with fallbacks) :

1. Daily note of the day : 00 - Daily notes/[target date].md

   • If absent → create minimal file (empty frontmatter + day title), continue
   • Format : ---\n---\n# YYYY-MM-DD\n

2. Sessions of the day : 99 - Claude Code/Sessions/YYYY-MM-DD.md

   • If absent → synthesize from daily note, note [Approx synthesis — to be validated by {USER_NAME}]
   • If multiple sessions → enumerate each with start time (ex: 09:30 — Project X)
   • If file too long (token error) → read by successive chunks (limit: 200, offset: 0 → 200 → 400 → …) until complete. Never build summary on partial read — if truncation, note [Complete read by chunks — synthesis validated].

3. Hobby Kanban : 02 - Hobbies/Hobby Kanban.md

   • If absent → skip Step 1.5 silently
   • Check if ## WIP column contains tickets (non-empty items)

4. Weekly lookback — 3 previous daily notes : 00 - Daily notes/YYYY-MM-DD.md for J-1, J-2, J-3

   • Targeted reading of ## 🌙 Day wrap-up, ## 💡 Ideas & Reflections, frontmatter (energy, score, feeling)
   • If a daily note is absent → skip silently, continue with available ones
   • Objective : detect inter-day patterns (energy rising/falling, recurring subjects, contradictions with today, repeated or isolated highlights)
   • These data feed Step 2.2 (sourced open questions) and Step 2.3 (general state narrative)

Step 2 — Build the wrap-up

2.1 — Aggregation and narrative grouping

From daily note and Sessions/YYYY-MM-DD.md, extract and group by project (not flat chronological order) :

• Grouping by project/theme : one feature, one project, one refinement session, one personal project = one bullet. Do not split if multiple sessions cover the same subject.
• Mandatory execution context per bullet : for each project, include what makes the work concrete — stack touched, files modified, tests written, ADR created, decisions made, people involved. No bare bullet like "Bug 429 fail to fetch" — prefer "Bug 429 + fail to fetch armed input/project — exponential retry on client side, Prisma timing debug, 3 tests added".
• Narrative highlights : explicitly call out marking events of the day (nice surprise, frustration, joy moment, notable interaction, crisis) — these go in General state at Step 2.3, not in the bullet list.
• Energy and states : observations of flow, fatigue, blockers, notable gaps (ex: low energy but high output) — for general state.
• Inter-day patterns : cross-referencing with 3 previous daily notes (loaded at Step 1, point 4), highlight what repeats, what contradicts, what emerges as constant. These observations feed open questions at Step 2.2.
• Semantic already-seen check (optional) : for each pattern identified above, query :

  uv run ~/.claude/semantic_search.py "<pattern in 3-5 words>" --top-k 2 --type completed
  

  If similarity ≥ 0.45 on a prior session → the pattern is a deja-vu. For each deja-vu :
  1. First appearance date : extract date from found session
  2. What came of it : relaunch with --type next_steps --since <date> to see planned follow-ups after this pattern
  3. Causal link : if follow-ups exist, build the link "pattern → what happened after" This information enriches the Patterns detected section in the note (see template below). If Postgres/Ollama inaccessible → non-blocking, display : ⚠️ Deja-vu verification unavailable (Postgres/Ollama down) — patterns sourced from daily notes only.
• Ideas/Decisions : flag but process in Step 4, not here.

2.2 — Presentation and validation of wrap-up

Present to {USER_NAME} the full narrative version — not a summarized version. {USER_NAME} must see in chat what will be written, not a flat preview :

Here's what I saw for the day :

What was done (grouped by project, context included) :

• [project 1 — execution context : stack, files, tests, ADR, decisions]
• [project 2 — execution context]
• …

General state (narrative, 2-4 sentences) : [energy felt, flow or friction, day highlight, possible link with fatigue/motivation — tell the day, don't summarize it]

Patterns observed (cross-referenced with 3 prior days) :

• [recurring pattern or contradiction detected, sourced "seen J-1/J-2" or "new vs yesterday"]
• [energy/output gap, or motivation compensating fatigue, etc.]

2 questions to dig deeper (sourced from weekly lookback, not generic) :

1. [Question linked to specific inter-day pattern, ex: "Pattern X came back 3 times this week — is it constant or context Y related ?"]
2. [Question on what notes don't capture well, or observed contradiction]

Rules for open questions :

• Always 1 to 2 questions, never zero (that's what makes the retrospective alive)
• Sourced : each question must rely on something concrete seen in the 3 last dailies or in today's thinking sessions
• Not generic : avoid "how do you feel?", "what did you learn?" — prefer precise questions on detected patterns or contradictions

If ## 💡 Ideas & Reflections exists and filled → present full narrative version + open questions, then ask : "Sound good, should I note it?"

• Accept : yes, ok, agreed, yeah. Reject : no, nope, silence/timeout (45s) → continue without modifying.

Otherwise (section empty or absent) → present what could be reconstructed + ask :

(1) What did you really advance ? (2) A blocker or observation ? (3) [question sourced from weekly lookback if patterns detected]

Accept responses (free form). Timeout (45s) or no response → include existing data (sessions, mood, weekly lookback) + note "[Quick Wrap skipped — timeout]".

2.3 — Writing the wrap-up

Check if ## 🌙 Day wrap-up already exists → If yes, skip to Step 3 (wrap-up already written). If no, add at bottom of daily note :

## 🌙 Day wrap-up

### What was done
- **[Project 1 — narrative title]** — complete execution context : stack touched, files modified, tests written, ADR created, decisions made, concrete results
- **[Project 2 — narrative title]** — complete execution context
- [...]

### Energy & state
- Energy of the day : [frontmatter value or "not filled"]
- Final score : [score]/5 — *[feeling]*
- **General state** : [NARRATIVE 2-4 sentences — tell the day, don't summarize it. Include : day's tone (atypical, dense, calm, difficult...), possible energy/output gap, flow moments, highlight(s) noted (good news, frustration, notable interaction, crisis, joy). No bullets here — text that tells a story.]

### Patterns detected
[Cross-reference with 3 previous daily notes. Integrate 2-4 structured observations :
- Inter-day patterns (recurrences, contradictions, energy/output gaps)
- Emerging behavioral dynamics sourced on concrete facts
- If `## 💡 Ideas & Reflections` is filled : do NOT just write "See Reflections section" — meta above it, highlight what emerges crossing multiple thinking sessions, not just list them
- For patterns identified as deja-vu (semantic verification Step 2.1) : mandatory format :
  `→ Already observed on YYYY-MM-DD — follow-up : [what came of it, in 1 sentence]`]

### Tomorrow
- Next priority step : [from sessions]
- Don't forget : [mentioned during the day]

Critical rule on narrative quality : this wrap-up is the main skill output. Steps 4 (ideas to pursue), 5 (capitalization proposals) and 6 (essay-check) are secondary and must in no way compromise wrap-up richness. If the model feels torn between "write the wrap-up well" and "anticipate lateral capitalizations", absolute priority to narrative wrap-up.

Step 3 — Score, feeling and frontmatter

Score — Check if score: exists in frontmatter (numeric value 1-5).

If present and valid → move to feeling.

If absent or invalid :

• Ask : "Rate your day on 5 ?"
• Accept : numbers 1-5. Reject : non-number, silence/timeout (30s) → skip without score.
• If obtained → write in frontmatter : score: [value]

(This score conditions the weekly essay-check at Step 6.)

Feeling — Always ask, whether score is already present or not :

• Ask : "One word on the day ?" (ex: "tough morning, good restart afternoon", "smooth throughout", "exhausting but satisfied")
• Accept : free text, 1 line max. Timeout (30s) → skip without feeling.
• If response → write in frontmatter : feeling: "[text]"

Step 4 — Ideas and decisions : filing suggestions

From sessions and daily note, flag all ideas/decisions mentioned : new insights, direction changes, tasks to track, tech to explore, etc.

If none found → skip silently, go to Step 5.

If found : propose a destination for each :

[Idea] → [Destination : 03 - Knowledge/ | 04 - Projects/… | 01 - Me/{USER_NAME}.md | project kanban | Hobby Kanban | command-tracker]

Present :

Ideas/decisions of the day :

• [item 1] → [destination]
• [item 2] → [destination]

Want to file some ?

Accept : explicit lists ("yes, the first 2", "item X only"). Reject : "no", "nope", silence/timeout (45s) → leave in daily note, write nothing.

Step 5 — Capitalization proposals (conditional)

Check if 99 - Claude Code/Sessions/proposals-YYYY-MM-DD.md exists for the day's date.

If absent → skip silently, go to Step 6.

If present and processed: true → skip silently, go to Step 6.

If present and processed: false :

1. Read the file
2. Present proposals to {USER_NAME} :

Capitalization proposals of the day :

[for each Session block] Session HH:MM :

• [ADR] Title — Scope — Context
• [Skill] Skill name — Action — Context

Do you validate, reject, or defer ?

3. For each validated item :
   • Transverse ADR → create in 99 - Claude Code/ADR/ + update ADR/INDEX.md
   • Project ADR → create in 04 - Projects/[Project]/claude-code/ADR/
   • Skill update → apply modification in 99 - Claude Code/Skills/[skill].md
   • Skill create → invoke /create-ticket with type Enhancement to ticket the creation
4. Timeout (45s) or global reject → skip all
5. Always mark processed: true in frontmatter after processing (validated or not)

Step 6 — Weekly essay-check (conditional)

Read 99 - Claude Code/command-tracker.md and find the last line - /essay-check → YYYY-MM-DD.

Conditions to trigger essay-check :

• command-tracker exists AND
• /essay-check entry found AND
• date in YYYY-MM-DD format valid AND
• date ≥ 7 days old (overdue)

Otherwise (file absent, entry not found, invalid format, or date < 7d) → skip essay-check, move to Step 7.

---

If essay-check overdue :

Trigger /essay-check and wait for return. (Full synthesis, check-in, and {USER_NAME}.md enrichment are handled by the dedicated essay-check skill.)

After essay-check : Update command-tracker : /essay-check → YYYY-MM-DD (today's date). Then move to Step 7.

Step 7 — Nudge measurement dreaming (conditional — Phase 0)

Read 99 - Claude Code/memory/digest.md.

Conditions to trigger nudge :

• file exists AND
• its frontmatter contains enabled: true

Otherwise (absent, enabled: false, or invalid format) → skip silently, end closeday. (The nudge thus self-disables once Phase 0 dreaming is turned off — no debt cleanup.)

---

If active :

Phase 0 of dreaming tests whether the working digest injected at SessionStart actually helps (cf. 99 - Claude Code/memory/phase0-log.md, spec §11 — "without metric, impossible to decide"). From today's sessions, propose to {USER_NAME} a line to add to the Journal table in phase0-log.md :

| YYYY-MM-DD | <session(s) of day, subject> | <✅ / ⚠️ partial / ❌ outdated> | <yes/no fetch avoided> | <yes/no hit useful> | <short note> |

• If nothing observable (digest not used, day outside vault) → propose a neutral line anyway (| … | — | — | — | no notable usage |) : the trace of absence counts for the verdict.
• No writing without {USER_NAME} validation — append at end of Journal table, never overwrite. Never judge "hit useful" on their behalf without confirmation.
• If {USER_NAME} flags that a live thread changed → propose updating digest.md + its updated: (optional, on demand).

Rules

1. Absolute priority to narrative wrap-up — the main skill output is a rich, narrative, project-grouped wrap-up, with weekly lookback and sourced open questions. Steps 4-5-6 (filing, capitalization, essay-check) are secondary and must never degrade wrap-up quality. If the model must arbitrate between narrative richness and lateral step anticipation, choose narrative richness.
2. No flat list — Step 2.3 : bullets in "What was done" must be grouped by project with complete execution context, never a flat chronological list. "General state" must be a 2-4 sentence narrative, not a bullet summary.
3. Always 1-2 sourced open questions at Step 2.2 — never zero, never generic. They must rely on concrete patterns observed in the 3 previous daily notes or today's thinking sessions.
4. No writing without confirmation — always present before writing
5. Standard timeouts — each interaction has a max (30-45s per context). On expiry : default action (skip, continue + note "[action skipped — timeout]")
6. Graceful missing files — skip silently if non-blocking (except today's daily note — create if absent)
7. Append-only — never overwrite daily note — ## 🌙 Day wrap-up sections added at bottom. If already exists → skip to Step 3.
8. Robust parsing — unexpected format (ex : score: "four", malformed date) → skip + note "[invalid format]", don't block
9. Chunking by slices — too long : limit: 200 (sessions) or 100 (mood), offset: 0 → 200/100 → … until end
10. {USER_NAME}.md fallback — if absent, skip all enrichment silently (Step 6 only)