Week in Review

A Friday ritual: you stayed in the flow all week — this skill surfaces what you were curious about so you can learn/deep-dive later, and shows how you spent your tokens, commands, skills, and ctx-stack usage.

Three deterministic scripts + one synthesis subagent. Do NOT reimplement the parsing by hand — the scripts already handle the Claude Code transcript format, cache-token accounting, and harness-wrapper stripping.

Paths (resolve these first)

# Where this skill is installed. Global install (default):
SKILL=~/.claude/skills/week-in-review
# ...or if installed as a PROJECT skill, use: SKILL="$(git rev-parse --show-toplevel)/.claude/skills/week-in-review"

# Where reports go. Override with WEEK_IN_REVIEW_OUT to point into a repo you sync/commit.
DAY=$(date +%F); OUT="${WEEK_IN_REVIEW_OUT:-$HOME/week-in-review}/$DAY"; mkdir -p "$OUT"

Inputs / defaults

• Window: defaults to this week (since Monday). Pick the flag that matches what the user asked for:
  • --last-week — the previous calendar week
  • --days N — a rolling last-N-days window ("the last 10 days")
  • --since YYYY-MM-DD [--until YYYY-MM-DD] — a specific date range or a named past week (omit --until for "since X until now"; a bare --until date counts through end of that day)
  • --week-start sun — treat the week as Sunday-started (applies to this-week / last-week)
  • When targeting a past range, name the output dir after it (e.g. .../week-in-review/2026-07-20_to_2026-07-26/) instead of today.
• Scope: all projects under ~/.claude/projects (override: --project-filter <substr>).
• Output: $OUT/report.html — a self-contained, theme-aware HTML file (its header shows the window it covers).

Steps

1. Collect the deterministic stats

python3 "$SKILL/scripts/collect.py" --out "$OUT/week-data.json"   # default: this week
# past week / range examples:
#   python3 "$SKILL/scripts/collect.py" --last-week --out "$OUT/week-data.json"
#   python3 "$SKILL/scripts/collect.py" --since 2026-07-20 --until 2026-07-26 --out "$OUT/week-data.json"

This prints a summary and writes week-data.json (token usage by model, slash commands, skills, ctx-stack subcommands, tools, projects, and your cleaned typed prompts). Read the printed summary so you know roughly what the week held.

2. Synthesise learning areas (subagent, live web search)

Spawn ONE subagent to do the clustering + link-finding — this is a search/clustering task, not frontier reasoning, so a mid-tier model (e.g. Sonnet) is the right call; don't burn the top session model on it. Give it this brief, substituting the real $OUT path:

Read $OUT/week-data.json. The prompts array is a week of the user's real typed prompts to Claude Code across their projects. Your job:

1. Cluster them into 3–8 coherent "learning areas" — topics the user asked about, explored, or seemed curious about. Ignore pure command/glue noise; focus on things worth learning more about later.
2. For each area write: a title, a one-line why_it_came_up, and 2–4 concise key_learnings bullets (what someone would take away).
3. For each area, run web searches to find 1–3 current, authoritative links (official docs, canonical references, high-quality articles). Prefer primary sources. Only include links whose URL you actually saw in search results — never invent URLs.
4. Also produce a short week_summary (1–2 sentences) and a curiosities list: things the user flagged wanting to explore later.

Write the result to $OUT/learnings.json with exactly this schema: { "week_summary": str, "areas": [ { "title": str, "why_it_came_up": str, "key_learnings": [str], "links": [ {"title": str, "url": str} ] } ], "curiosities": [str] } Output only valid JSON to the file. Confirm the path when done.

If the subagent cannot search (offline), it should still cluster and distil, omitting links — the report degrades gracefully.

3. Render the HTML report

python3 "$SKILL/scripts/render.py" \
  --data "$OUT/week-data.json" --learnings "$OUT/learnings.json" --out "$OUT/report.html"

render.py is defensive: if learnings.json is missing or partial it still renders the usage dashboard plus a raw-prompt fallback, so always run it.

4. Hand off

Tell the user the path ($OUT/report.html) — a self-contained, theme-aware file they can open in any browser — with a one-line highlight from the summary.

Guardrails

• The report quotes your prompts — treat it as personal. By default it lives outside any repo; don't publish it to an external/hosted service unless asked.
• Token totals are dominated by cache_read (context re-served from cache) — the report deliberately separates fresh Output/Input from cache reads so the numbers aren't misleading. Don't "correct" this by summing everything.
• "ctx-stack usage" refers to the ctx CLI from the AI Codebase Semantics toolkit; on machines without it that section is simply empty.