Emotional Recap

Generate a wellbeing-focused review of the user's recent Claude Code conversations. The skill reads conversation transcripts, performs sentiment / tone / emotion analysis using established psychological frameworks, surfaces patterns (time-of-day, project, streaks, vocabulary), and writes a supportive Markdown + HTML report grounded in cited research.

This skill is supportive, not diagnostic. It surfaces observations and reframes them with science; it never pathologises, prescribes treatment, or replaces a professional. Wellness is a long game — this is one signal among many.

Preferences

_On startup, use the Read tool to load ~/.claude/skills/emotional-recap/preferences.md. If missing, treat as "no preferences set" and apply Defaults from this file._

Design contract

_On startup, use the Read tool to load ~/.claude/skills/emotional-recap/DESIGN.md. This file is the renderer contract: aesthetic direction, type system, colour tokens, layout, visualisation principles, anti-patterns, renderer guarantees, component patterns. Apply its tokens and rules to every emitted HTML / SVG / Markdown artefact. Renderer guarantees in DESIGN.md §7 are non-negotiable — they fail closed (refuse to write if violated)._

_When DESIGN.md and preferences.md conflict: DESIGN.md wins for renderer guarantees (a11y, anti-patterns, token shape); `preferences.md` wins for user-tunable knobs (tone, citation depth, framework choice, auto_open, etc.). If DESIGN.md is missing, fall back to the inline spec in reference/report-structure.md and flag once._

Context

_On startup, use Bash to detect: current working directory, current project slug (cwd path with / → -, prefixed with -), today's date in YYYY-MM-DD. Skip any that fail._

Conversation transcripts live at ~/.claude/projects/<project-slug>/<uuid>.jsonl. Each line is one event (user message, assistant message, tool result, system reminder). The skill operates on user-authored text only — assistant output, tool outputs, and system reminders are excluded from sentiment analysis (they are not the user's emotional signal).

Command routing

Check $ARGUMENTS:

• `help` → display Help block, stop
• `config` → interactive setup, stop
• `reset` → delete preferences.md, feedback-journal.md, sessions/, resume-state.md, reports/, confirm, stop
• `feedback` → run Feedback subcommand (see ## Feedback & learning)
• `resume` → if resume-state.md exists, continue the previous unfinished run
• `--window=<dur>` → override lookback (e.g. 24h, 7d, 30d)
• `--scope=<all|current|slug>` → override conversation scope
• `--format=<both|md|html>` → override report format
• `--last=N` → limit to N most recent conversations (after window filter)
• `--dry-run` → analyse + print summary only, skip writing report
• anything else / empty → run the main workflow

Help

emotional-recap — Wellbeing review of recent Claude Code conversations

Usage:
  /emotional-recap                            Default: all projects, last 48h, both formats
  /emotional-recap --window=7d                Last 7 days
  /emotional-recap --scope=current            Only the current project's conversations
  /emotional-recap --format=html              HTML report only
  /emotional-recap --last=10                  Last 10 conversations (after window filter)
  /emotional-recap --dry-run                  Analyse + print summary, no file written
  /emotional-recap feedback                   Rate the most recent report
  /emotional-recap config                     Set preferences
  /emotional-recap resume                     Continue an interrupted run
  /emotional-recap reset                      Clear preferences + journal + reports
  /emotional-recap help                       This help

Current preferences:
  (loaded from ~/.claude/skills/emotional-recap/preferences.md)

Config

Use AskUserQuestion to collect (in two batches of ≤4 questions):

Batch 1 — Scope & output

• Q1 — Default lookback window: 24h / 48h / 7d / 30d
• Q2 — Default scope: all projects / current project only / ask each time
• Q3 — Default format: both md+html / md only / html only
• Q4 — Report directory: ~/.claude/skills/emotional-recap/reports/ (default) / ~/Desktop/wellness/ / custom path

Batch 2 — Tone & depth

• Q5 — Auto-open HTML in browser after generation: yes / no
• Q6 — Tone of voice: warm & supportive (default) / clinical & neutral / direct & terse
• Q7 — Citation depth: minimal (1-line refs) / standard (full APA inline, ~5 refs) / academic (full reference list, ~10 refs)
• Q8 — Primary emotion framework: Plutchik wheel (8 primary emotions) / Ekman basic (6 emotions) / Russell circumplex (valence×arousal) / combined (use all three contextually)

Write to ~/.claude/skills/emotional-recap/preferences.md in three-tier format:

# /emotional-recap preferences
Updated: YYYY-MM-DD

## Defaults
- window: 48h
- scope: all
- format: both
- report_dir: ~/.claude/skills/emotional-recap/reports/
- auto_open: yes
- tone: warm
- citation_depth: standard
- framework: combined

## Profile (optional — edit freely)
- (free-form lines the user can edit to nudge defaults — e.g. "avoid mentioning sleep; I work nights by choice")

## Learned
- (populated from feedback over time)

Reset

Delete in order, with confirmation:

1. ~/.claude/skills/emotional-recap/preferences.md
2. ~/.claude/skills/emotional-recap/feedback-journal.md
3. ~/.claude/skills/emotional-recap/sessions/
4. ~/.claude/skills/emotional-recap/resume-state.md
5. Ask before deleting ~/.claude/skills/emotional-recap/reports/ — these are user-facing artefacts.

Confirm: Preferences, journal, sessions, and resume state cleared. Reports kept unless you opted in. Using defaults.

First-time detection

If ~/.claude/skills/emotional-recap/preferences.md does not exist:

First time using /emotional-recap? I'll use sensible defaults (last 48h across all projects, both md+html, warm tone, combined emotion frameworks). Run /emotional-recap config anytime to tune. Continuing…

Then proceed.

Workflow

Step 0 — Load learning context and design contract

1. Read ~/.claude/skills/emotional-recap/preferences.md → apply Defaults, then Profile overrides, then Learned overrides.
2. Read `~/.claude/skills/emotional-recap/DESIGN.md` — the renderer contract. Hold its tokens, anti-patterns, and renderer guarantees in working memory for Step 5. If missing, fall back to reference/report-structure.md and flag once.
3. Read ~/.claude/skills/emotional-recap/feedback-journal.md if present — scan the last ~10 entries for Signal: lines and silently apply them as soft biases for this run.
4. If resume-state.md exists and $ARGUMENTS is not resume, mention once: A previous run was interrupted — run /emotional-recap resume to continue, or ignore to start fresh. Continue with the new run regardless.

Continue silently on missing files — graceful degradation is mandatory.

Step 1 — Resolve scope and window

1. Parse flags from $ARGUMENTS; fall back to preferences; fall back to Defaults.
2. Compute window cutoff: now - window (e.g. now - 48h).
3. Determine projects to scan:

• scope=all → glob ~/.claude/projects/*/
• scope=current → only the current project slug
• scope=<explicit-slug> → that one directory

1. List JSONL files inside each project directory, filter by mtime >= cutoff.
2. If --last=N, sort by mtime desc and take top N.
3. If zero files match → print a warm note and stop. Example: No conversations found in the last <window>. Either nothing's happened, or you're not using Claude Code through it — enjoy the quiet.

Print a one-line confirmation: Scanning N conversations across M projects from <date> to <date>.

Step 2 — Extract user-authored text

For each JSONL file:

1. Read line by line. Each line is JSON.
2. Keep only events where the author is the user (typically type == "user" with a message.content that is a string or has text parts). Exclude:

• Assistant messages
• Tool results / tool outputs (e.g. tool_use_result, tool_result)
• System reminders and <command-*> tag payloads
• Slash command bodies (lines starting with <command-name> blocks) — keep only the natural-language fragments the user typed
• Empty / whitespace-only messages

1. Capture per message: timestamp, project_slug, conversation_id, text.
2. Skip files that fail to parse — log a soft warning but continue.

Aggregate stats: total messages, total tokens (approx word count × 1.3), distinct conversations, distinct projects, span (first → last timestamp).

Step 3 — Sentiment, emotion, and tone analysis

Apply the user's chosen framework(s) — see ~/.claude/skills/emotional-recap/reference/frameworks.md for the full reference. Summary:

• Plutchik (1980) — 8 primary emotions: joy, trust, fear, surprise, sadness, disgust, anger, anticipation. Tag each meaningful user message with 0-2 primary emotions.
• Ekman (1992) — 6 basic emotions: happiness, sadness, fear, anger, surprise, disgust. Useful when Plutchik feels too granular.
• Russell circumplex (1980) — every message gets a valence (−1 unpleasant → +1 pleasant) and arousal (0 calm → 1 activated) score.
• Pennebaker LIWC-style markers (2003) — track linguistic categories: 1st-person pronouns (self-focus), negative emotion words, cognitive complexity ("because", "however"), certainty markers ("never", "always"), social references.

For each message produce a compact record:

{ts, project, valence, arousal, plutchik: [...], dominant_tone, markers: {...}}

Important reading caveats (include in every report):

• This is text-only analysis. It misses sarcasm, context, and intent. Treat it as a mirror, not a verdict.
• Short technical messages ("fix the bug", "ok") are tone-neutral, not negative. Don't over-read.
• Frustration in debugging is normal cognitive work, not distress — see Csikszentmihalyi's flow (1990): productive struggle has the same surface markers as suffering.

Step 4 — Pattern detection

Surface the patterns that actually matter. Look for:

1. Temporal arc — day-by-day mean valence + arousal. Highlight the trend (rising / falling / flat) and the variance (stable / volatile).
2. Time-of-day — bucket into morning (06-12), afternoon (12-18), evening (18-22), late-night (22-06). Flag if late-night dominates AND valence drops there.
3. Project correlation — mean valence per project. Flag if one project is >0.4 below the user's overall mean (a meaningful gap, not noise).
4. Streaks — 3+ consecutive sessions in the same emotional quadrant (Russell circumplex). E.g. "3 sessions in high-arousal / low-valence — frustration cluster."
5. Vocabulary shifts — compare this window's top emotional words against the journal's baseline (if learned). New words = new state.
6. Self-focus drift — Pennebaker: rising 1st-person pronoun frequency correlates with rumination / depressive ideation. Flag only if change >50% from baseline, never on first run.
7. Cognitive complexity — falling "because/however/although" use can indicate fatigue or stress narrowing thinking.

Pattern presentation rule: every flagged pattern must include (a) the observation, (b) the science behind why it matters, (c) one supportive reframe or question — never an instruction. Example:

_Your messages got shorter and more 1st-person after 11pm on May 11 and 12. Pennebaker (2003) found rising self-focus often tracks with fatigue or rumination. Worth noticing — what was happening those nights?_

Step 5 — Generate the report

Render both formats from a single shared structure. Bind every visual decision to a `DESIGN.md` token — no hard-coded hex, no inline style="" for colour or layout (data-driven values like width: 60% for a bar fill are allowed). Use the component patterns in DESIGN.md §8 and the renderer rules in reference/report-structure.md.

Before writing the file, run the 16-item renderer self-audit (DESIGN.md §7 → "Renderer self-audit"):

1. Every SVG has role="img", <title>, <desc>.
2. Every decorative span has aria-hidden="true".
3. No color-mix() in emitted CSS — pre-compute mixes for both themes and emit as theme-scoped CSS vars (see item 15).
4. No title=""-only tooltips on interactive elements.
5. Russell scatter passes the overplotting guard (jitter applied; density-bin mode triggered if needed).
6. Anonymisation pipeline ran: deny-list + heuristic + structural strip. No client / brand names in quotes.
7. Section numbering 01–12 contiguous (no gaps).
8. Hero number paired with −1 negative · 0 neutral · +1 positive legend.
9. All text-bearing colour tokens used clear AA contrast against their background.
10. SVG fills bound to var(--*) tokens, not literal hex.
11. Sparse-data fallbacks emit "Not enough data" text inside affected SVG regions instead of misleading charts.
12. Touch-target sizes ≥44px on (pointer: coarse) viewports for interactive chart elements.
13. Chart provenance — every chart-bearing section (§05–§10) emits a <p class="chart-source"> caption with framework/method + a real citation drawn from reference/citations.md. Per DESIGN.md §7 provenance table.
14. Worth-trying integrity — if §04 is non-empty: 2–4 items, each with one citation from reference/citations.md, no prescriptive phrases ("you should", "you must", "fix", "stop", "treatment", "therapy"). If §04 is empty, the section header still renders with the "no standout options this window" line; do not silently skip.
15. Theme-parity (1225 bug fix) — grep the emitted CSS for hex literals (#[0-9a-fA-F]{3,8}) appearing as background, color, or border-color values on selectors that hold text or visible surface. Allowed locations for hex literals: inside :root { … } and @media (prefers-color-scheme: dark) :root { … } blocks ONLY. Anywhere else → fail.
16. Per-theme contrast check — for every text-on-surface pairing (e.g. .hl-quote color on .hl-card background), compute the contrast ratio using the light-mode token values and again using the dark-mode token values. Fail if either ratio is < 4.5:1 for normal text or < 3:1 for large text (≥18pt or ≥14pt bold).

If any item fails, refuse to write and report which check failed. The user may override with --force but the default is fail-closed.

The audit is not a substitute for visual inspection. If a render passes 16/16 but still looks wrong in the reader's theme, the audit needs another item — log a journal entry and grow the checklist. Quality is the union of "passes the rules" + "looks right in the actual theme the reader uses."

Section order (DESIGN.md §4 IA — _fold-by-fold_: brief → highlights → insights → anchors → raw data → reference; 12 sections numbered 01–12 contiguously; both md + html unless noted):

_— Brief —_

1. At a glance — 2-3 sentence brief + hero valence with scale legend + quadrant chip + 4 stat tiles + daily badge row. The whole top-fold story.

_— Highlights —_

1. Highlights — compact grid of 3–5 bright-spot chip-cards (anonymised quote + date). Promoted from bottom to top fold per 2026-05-13-1150 feedback. Required (Fredrickson, 2001).

_— Insights —_

1. Insights — horizontal carousel of pattern cards (observation / science as <details> / reframe). 3–7 cards. Scroll-snap, no JS, accessible in source order. This is the report's middle act — the main signal.

_— Anchors —_

1. Worth trying — numbered list of 2–4 evidence-informed micro-experiments, each anchored to a real citation from reference/citations.md. Never prescriptive — the language is "research suggests…" / "studies indicate…" / "worth trying if it fits". Per DESIGN.md §6 language guardrails. Section may render empty (with the "no standout options" line) if no pattern this run admits a research-backed action.

_— Raw data — every chart carries a chart-source caption with framework + citation —_

1. Emotional arc — twin sparklines (valence + arousal) + arc narrative.
2. Affect map — _(HTML)_ Russell scatter with jitter + density-bin fallback + per-quadrant tint + size variation. _(MD)_ ASCII quadrant.
3. Dominant emotions — _(HTML)_ Plutchik wheel + size-redundant outer dot ring + ranked bars + quoted examples. _(MD)_ ranked list with %.
4. Day rhythm — 24-hour focusable strip with peak/dip callouts and sample-size warnings.
5. By project — diverging horizontal bars centred at zero; opacity scaled by message count.
6. Word weather — typographic vocabulary block (Fraunces italic). Polarity = colour + weight + underline-on-negative + aria-label.

_— Reference —_

1. Citations — bibliographic list per the user's citation_depth setting. Pull from reference/citations.md.
2. Caveats — text-analysis caveats from Step 3 + the "this is not clinical advice" line.

HTML design requirements — see DESIGN.md for the canonical contract (tokens, type, palette, anti-patterns, renderer guarantees). The high-level shape:

• Visual-first. Every section (except 08, 10, 11) leads with a chart, SVG, or typographic visualisation. <10-second scan target.
• Type system per DESIGN.md §2: Fraunces (display) + IBM Plex Sans (body) + IBM Plex Mono (mono). No Inter. No -webkit-font-smoothing: antialiased. No font-feature-settings: "ss01" on fallback fonts.
• One display number per report — the hero valence at --t-hero (5.5rem). Always paired with −1 negative · 0 neutral · +1 positive legend below.
• Tabular figures (font-feature-settings: "tnum") on every numeric value.
• Section numbering 01–11 contiguous with .num span marked aria-hidden="true". No gaps.
• Quadrant palette per DESIGN.md §3 — each chart binds to its appropriate --q?-line / --q?-ink / --q?-fill tokens. AA-compliant text tokens only.
• `color-mix()` forbidden in emitted CSS — renderer pre-computes mixes to hex.
• Inline SVG only — no external assets. Every SVG has role="img", <title>, <desc>.
• Interactive chart elements are <g tabindex="0"> with focus-visible state + inline-revealed details. No title=""-only tooltips.
• **Single <style> block in <head>. Dark mode via prefers-color-scheme: dark with mirror token set.
• Max-width 720px, hero allowed one asymmetric left-bleed on viewports ≥820px (DESIGN.md §4).
• Anonymisation pipeline mandatory per DESIGN.md §7 — three-layer (deny-list → heuristic → structural). No client / brand names in quotes.

Tone of voice: per preferences (warm default).

• Warm: "I noticed your evenings have been heavier this week — your mornings sounded brighter. That kind of swing is common when sleep gets short (Walker, 2017)."
• Clinical: "Mean valence in evening sessions: −0.18; mornings: +0.21. Diurnal mood variation correlates with sleep architecture (Walker, 2017)."
• Direct: "Evenings are dragging. Mornings are fine. Sleep matters here — see Walker (2017)."

File naming: reports/YYYY-MM-DD-HHMM-recap.md and .html.

Step 6 — Save, log, and (optionally) open

1. mkdir -p the report directory if needed.
2. Write both files (per format preference).
3. Append to ~/.claude/skills/emotional-recap/sessions/YYYY-MM-DD-HHMM.md:

   # Session YYYY-MM-DD HH:MM
   - window: <value>
   - scope: <value>
   - conversations: N
   - dominant_quadrant: <Russell quadrant>
   - mean_valence: <x>
   - mean_arousal: <y>
   - top_emotions: [...]
   - patterns_flagged: [<short list>]
   - report_md: <path>
   - report_html: <path>

1. If auto_open is true and the format includes html, run open <path> (macOS) / xdg-open (Linux). Skip silently on failure.

Step 7 — Summarise to the user

Print, in chat (≤6 lines):

• One-line vibe summary
• Top 2 patterns
• File paths

Example:

Window: last 48h, 14 conversations across 3 projects. Mean tone: slight positive, moderately activated. 2 patterns to consider: (1) evening valence drop, (2) frustration streak in azad repo (3 sessions). Saved: ~/.claude/skills/emotional-recap/reports/2026-05-13-1430-recap.md (+ .html)

Step 8 — Invite feedback

Final line:

Run /emotional-recap feedback — even one rating sharpens the next report. This is a mirror, not a diagnosis — your felt sense always overrides the analysis.

Feedback & learning

When invoked as /emotional-recap feedback:

1. Find the most recent sessions/YYYY-MM-DD-HHMM.md. If none → say No recent session found. and stop.
2. Print a one-line summary of that session (vibe + top pattern).
3. Ask via AskUserQuestion (4 questions, one batch):

• Q1 — Tone fit: warm / clinical / direct / preachy / off
• Q2 — Reading accuracy: matched my experience / mostly right / off-base / can't tell
• Q3 — Pattern usefulness: insightful / generic / missed the real thing / too many flags
• Q4 — HTML design: loved / fine / cluttered / didn't open

1. Append to ~/.claude/skills/emotional-recap/feedback-journal.md:

   ## YYYY-MM-DD-HHMM
   - tone_fit: <answer>
   - accuracy: <answer>
   - usefulness: <answer>
   - design: <answer>
   - Signal: <one-line generalisation, e.g. "user finds 'warm' tone preachy; bias to 'clinical' next run">

1. Promotion rule: when 3+ sessions show the same Signal, promote it to ## Learned in preferences.md. Mention once: Noticed you consistently <signal>. Saved as a standing default — change anytime in /emotional-recap config.
2. Drift correction: when a Learned rule is contradicted in 2 newer sessions, demote it (move out of ## Learned, log the demotion in the journal with a Demoted: line). Never leave stale rules.

Pre-selection bias: in Step 5 of the main workflow, the tone-of-voice choice and pattern-density threshold are biased by current ## Learned rules. E.g. if Learned says "user wants ≤3 flags per report", cap pattern flags at 3.

Resume

If ~/.claude/skills/emotional-recap/resume-state.md exists and $ARGUMENTS == resume:

1. Read the state file. Expected keys: step, window, scope, files_processed, intermediate_path.
2. Resume from the noted step using the saved scope. Skip files already processed.
3. On success, delete resume-state.md.

The main workflow writes resume-state.md only when it processes ≥10 files — small runs are not worth resuming. State is overwritten on each run.

Principles

1. Supportive, not diagnostic. This skill is a mirror, not a medical tool. Every observation has a science citation and a reframing question, not a prescription. Mental health concerns belong with a professional.
2. Graceful degradation. Missing preferences.md / feedback-journal.md / failed JSONL parse — continue with defaults. Never block.
3. User text only. Only user-authored messages are analysed. Tool outputs and assistant messages are not the user's emotional signal.
4. Cite the science. Every pattern flagged must include a referenced finding. No folk psychology, no horoscope language.
5. Bright spots are required. Wellbeing reports that only flag problems become a negativity loop. Always include positive markers.
6. Learn quietly. Promote a Learned rule only after 3+ consistent signals. Mention once when promoting.
7. Drift correction over staleness. A Learned rule contradicted twice gets demoted. Never carry rules that no longer fit.
8. Destructive actions need confirmation. reset always asks before deleting reports/ (they are user artefacts, not skill state).
9. Self-awareness over self-judgement. The point is noticing patterns, not scoring the user. Frame everything as observation + curiosity.
10. Stop means stop. If the user halts mid-run on a ≥10-file analysis, save resume state. Don't silently keep working.

Self-audit notes

• Three-tier preferences: ✅ Defaults / Profile / Learned in Config block.
• feedback, resume, config, reset, help subcommands: ✅ all routed.
• Session logs + feedback journal + promotion + drift correction: ✅ in Feedback & learning section.
• Pre-selection bias from Learned: ✅ called out in Step 5.
• Configurability knobs (≥3): ✅ window, scope, format, report_dir, auto_open, tone, citation_depth, framework (8 knobs).
• Reference material externalised: reference/frameworks.md, reference/citations.md, reference/report-structure.md to keep this file under 500 lines.