weekly-review - SKILL.md Agent Skill

name: weekly-review description: >- Generate a weekly research digest — activity summary, knowledge growth, priorities. TRIGGER when: user asks for a weekly review, research summary, progress check, or "what did I do this week." Reads events.jsonl, research-state.yaml, wiki, mental-gym status, and IDEAS.md to produce a structured report. allowed-tools: Agent, Read, Glob, Grep, Edit, Write, Bash, WebSearch, WebFetch argument-hint: [optional: "last N days" or date range]

Weekly Review — Research Digest

You produce a structured research digest that helps the researcher see their trajectory, identify gaps, and set priorities. This is reflection and planning, not just reporting.

ultrathink

Context Loading

Read all sources in parallel:

events.jsonl — filter to the review period (default: last 7 days). Count events by type.
.claude/research-state.yaml — current gym, wiki, and session state
IDEAS.md — check for changes (git diff if available), note new items
Wiki pages — Glob all pages, check last_reviewed dates, count by type. Include wiki/research-evaluations/ in the counts (these are persisted verdicts from /research-companion).
Mental Gym status — run cd mental-gym && .venv/bin/mental-gym status to get current mastery data
Git log — run git log --oneline --since="7 days ago" for commit activity
.review/ directory — check for unresolved academic review findings

If a date range is provided in $ARGUMENTS (e.g., "last 14 days", "March 1-7"), adjust the filter period accordingly. Default is 7 days from today.

Report Structure

Present the digest in this format:

Weekly Research Digest — [Start Date] to [End Date]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Activity Summary
  Papers ingested: [N] ([list titles if any])
  Training sessions: [N] ([total exercises], avg score [X])
  Wiki pages updated: [N]
  Experiments run: [N]
  Writing reviews: [N] ([total findings])
  IDEAS.md updates: [N]
  Commits: [N]

Knowledge Growth
  Wiki: [total pages] pages ([+N new] this week, includes research-evaluations/)
  Topics mastered (>70%): [list or "none yet"]
  Topics improved: [list with deltas, e.g., "validation-levels: 0% → 35%"]
  Topics declining: [list with days since last review]
  New concepts added: [list]

Decisions made this week
  [list each research-evaluations/ page touched this week with its verdict,
   e.g. "agent-populations-for-polling — PURSUE",
        "simulators-as-survey-instruments — PARK (revisit: Q3 data release)"]
  [omit this block if empty]

Research Trajectory
  Active threads: [from IDEAS.md and recent events]
  Decisions made: [any PURSUE/PARK/KILL from research-companion]
  Dots connected: [any new connections identified]
  Questions answered: [any wiki queries promoted]

Health Checks
  Stale wiki pages (>60 days): [list]
  Orphan wiki pages: [list]
  Unresolved review findings: [count and file]
  Training streak: [N days, or "broken — last session [date]"]
  IDEAS.md last updated: [date]

Analysis Layer

After the report, add a brief analysis section (3-5 bullet points):

Patterns

What topics are getting the most attention? Is this aligned with priorities?
Are there blind spots — topics that haven't been touched?
Is the read→ingest→train flywheel running, or is one step being skipped?

Connections

Did any papers ingested this week connect to each other or to existing work?
Are there IDEAS.md dots-to-connect that now have more evidence?
Did training reveal knowledge gaps that should drive the next reading?

Risks

Is any project stuck or losing momentum?
Are there stale wiki pages in areas of active research?
Is the researcher learning broadly but not deeply, or vice versa?

Parked ideas ready to revisit (soft heuristic)

Scan wiki/research-evaluations/*.md for any PARK verdict where:

revisit_conditions in frontmatter is non-empty, AND
the evaluation date is older than 30 days before today, AND
the topic (or a close keyword) appears in events from the review period (e.g. a paper ingestion, IDEAS.md update, or wiki edit that touched the same area).

Flag these gently — "PARKed idea '[topic]' may be worth revisiting: condition was '[revisit_conditions]', and [event] this week touches that area." Do not auto-promote; this is a nudge. Skip silently if wiki/research-evaluations/ doesn't exist.

Priorities for Next Week

Based on the analysis, suggest 3-5 priorities ranked by impact:

Priorities for [Next Week Dates]
  1. [Priority] — [why it matters] — [specific action]
  2. ...
  3. ...

Priority categories:

Knowledge gaps: Topics with low mastery that are needed for active work
Stale knowledge: Wiki pages or training topics that need refreshing
Momentum: Projects that need a push to stay on track
Opportunities: New connections or ideas worth exploring
Maintenance: Wiki lint, review findings, broken streaks

Optional Actions

After presenting the digest, offer:

Update IDEAS.md — "Based on this week's activity, I'd suggest adding/updating these items in IDEAS.md: [list]. Want me to make these edits?"
Graduate mature ideas — "These IDEAS.md items seem ready to become wiki pages: [list]. Want to promote them?"
Wiki lint — "There are [N] health issues in the wiki. Want me to run a full lint and fix them?"
Set focus for next week — "Want to set a specific focus for next week? I'll adjust your research-state.yaml suggested actions."
Drafts touched but not reviewed — Cross-reference events.jsonl writing:edit events in the review period against .review/ findings. For each draft file that has writing:edit events but no corresponding entry under .review/ (or a stale one older than the most recent edit), list it and suggest /academic review <file>. Skip this pass silently if the /academic skill is not installed (check ~/.claude/skills/academic/SKILL.md or a plugin equivalent); the suggestion line should also be omitted in that case.
Share as event — Emit a weekly review event:

{"ts":"...","type":"review:weekly","detail":"Week of [dates]: [N] papers, [N] sessions, [N] wiki updates","source":"weekly-review"}

Orchestration Rules

Be honest about low activity. If nothing happened this week, say so without judgment. "Quiet week — that's fine. Here's what might be worth picking up."
Spot patterns the researcher might miss. You see the data across all tools — use that holistic view.
Don't moralize about streaks. Report them factually. Some weeks are for thinking, not training.
Suggest, don't prescribe. Present priorities as suggestions. The researcher knows their context better.
Keep it scannable. The digest should be readable in 2 minutes. Use the structured format above.

User's Input

$ARGUMENTS