fractal-review

The human’s reflective practice, AI-facilitated. Kepano’s fractal journaling: capture fragments throughout the day, compile them into reviews at increasing timescales, trace how ideas emerged and bubbled up into bigger themes. The AI gathers, surfaces, and prepares — the human does the actual reflecting. The maintenance is the understanding. Don’t delegate understanding.

Philosophy

Fractal journaling is not summarization. The AI’s job is to make it easy for the human to reflect, not to reflect for them. The human’s voice is the signal; the AI is the librarian who pulls the right files and says “you might want to look at these.”

Three principles:

1. Prepare the surface, don’t write the review. Gather fragments, surface themes, highlight connections — then hand the pen to the human.
2. Cascading timescales. Daily fragments feed weekly reviews. Weekly reviews feed monthly. Monthly feed yearly. Each level distills, never duplicates.
3. The human writes the review. AI creates created-by: ai preparation notes. The review itself is created-by: human. The human’s words are the durable artifact.

When to use

• Heartbeat detects a review is due (see cadence table below)
• “Review my week”, “monthly review”, “what happened today/this week/this month”
• “Prepare my yearly review”, “help me reflect”
• End of day when fragments exist
• Any time the human wants to look back

Inputs

Cadences

Process

STEP 1 — Detect cadence and gather material

If cadence isn’t specified, check what’s due:

1. Read the most recent review at each cadence. Check daily/ for the daily note, vault root (or personal/journal/) for weekly/monthly/yearly reviews. Use filename patterns: YYYY-[W]ww.md, YYYY-MM.md, YYYY.md.
2. Determine which reviews are overdue. A weekly review is due if >7 days since the last one. Monthly if >30 days. Yearly if >365 days. Start with the smallest overdue cadence.
3. If multiple cadences are due, work bottom-up. Do daily → weekly → monthly → yearly in sequence. Each feeds the next.

STEP 2 — Gather fragments for the period

Based on the cadence, collect the raw material. The Daily.base (embedded on each daily note) provides ready-made views: Everything (default — all notes except pure AI-generated, including notes without metadata), Human (only created-by: human), Fragments (journal fragments for that date), Reviews (weekly/monthly/yearly reviews covering that date), AI (only created-by: ai or ai-assisted). Use these as a starting point — they do most of the gathering for you.

Daily:

• The Daily.base → Everything view shows all notes linked to today except pure AI notes (the default tab)
• The Daily.base → Fragments view shows notes with the YYYY-MM-DD HHmm prefix pattern
• Find all notes modified today (check file.mtime)
• Find all links made to today’s daily note ([[YYYY-MM-DD]] backlinks)
• List any notes the human touched, bookmarks processed, tasks completed

Weekly:

• Find all fragments from the past 7 days
• Find all daily notes from the week (check backlinks on each)
• Collect: new notes created, notes promoted (fleeting → permanent), connections made, tasks shipped, bookmarks processed
• Surface: recurring themes (concepts that appeared 3+ times), open questions, unresolved threads

Monthly:

• Read all weekly reviews from the month
• Collect: new knowledge notes, meta observations, goal progress, project movement
• Surface: patterns across weeks, shifting interests, emerging dimensions

Yearly:

• Read all monthly reviews from the year
• Collect: major milestones, biggest insights, relationships deepened, projects shipped/abandoned
• Surface: growth arc, identity shifts, recurring struggles, breakthrough moments

5-year:

• Read all yearly reviews in the range
• Surface: life trajectory, identity evolution, decade themes, what you’d tell your past self

STEP 3 — Prepare the review surface

Create a preparation note: inbox/reviews/YYYY-MM-DD Review prep — [cadence].md

---
type: log
created: YYYY-MM-DD
created-by: ai
tags: [review-prep]
---

The prep note contains:

1. Period summary — what timeframe this covers, how many fragments/notes/reviews exist
2. Fragment digest — each fragment with its opening line and links, grouped by day (for weekly) or week (for monthly)
3. Themes detected — recurring concepts, people, places, emotions. Link each to vault notes.
4. Connections surfaced — links the human made during this period that bridge previously unconnected ideas
5. Open threads — questions asked but not answered, tasks started but not finished, ideas seeded but not developed
6. Random revisit candidates — 3–5 older notes (from before this period) that connect to this period’s themes. The human may want to revisit them.
7. Prompt questions — 3–5 reflective questions tailored to what actually happened (not generic). E.g. “You mentioned [[burnout]] twice this week but also shipped [[Project X]]. What’s the relationship?”

For yearly reviews: Include the 40 questions template as a starting scaffold.

STEP 4 — Invite the human to review

Present the prep note to the human in chat. Don’t summarize their life — show them the raw material and ask:

“You have [N] fragments from this [week/month/year]. I’ve prepared a review surface with themes I noticed and some questions. Ready to write your [weekly/monthly/yearly] review? I’ll create the note and you fill it in.”

If the human says yes, create the review note from the appropriate template (see below). The note is created-by: human — the human writes in it.

STEP 5 — Support the review (if the human engages)

While the human is writing their review:

• Answer questions about what happened (“what did I write about on Tuesday?”)
• Surface specific fragments on request (“show me everything about [[Project X]] this month”)
• Suggest connections (“your note about [[burnout]] connects to what you wrote in [[2026-W03]]”)
• Do not write the review for them. Reflect back, surface, connect — but the words are theirs.

STEP 6 — Post-review linking

After the human finishes their review:

1. Link the review to the period’s daily notes. Add [[YYYY-[W]ww]] links to each daily note in the period (as backlinks — the daily notes stay empty, but the review now appears in their backlinks pane).
2. Link upward. If a monthly review exists, link this weekly review to it. If a yearly exists, link the monthly.
3. Extract durable insights. If the human wrote something that deserves its own atomic note (a realization, a decision, a pattern), suggest it. Create the note only with their approval — it will be created-by: ai-assisted since the insight came from the human’s review.
4. Update meta. If the review surfaces a preference, alignment insight, or reasoning pattern, note it in meta/ and update AGENTS.md if it changes a convention.

Review note naming

Location follows the human’s preference. Default to wherever they’ve been writing journal fragments. If no preference established, ask.

Outputs

• Review prep note (inbox/reviews/, created-by: ai) — the AI’s gathered material
• Review note (created-by: human) — the human’s actual reflection
• Updated daily note backlinks connecting the period
• 0+ atomic notes extracted from the review (created-by: ai-assisted)
• Agent log entry (inbox/log/YYYY-MM-DD) summarizing which review was facilitated

Decision authority

• YOU DECIDE: what to gather, which themes to surface, prompt questions, connection suggestions, which older notes to recommend for revisit
• ESCALATE TO HUMAN: the review itself (always), extracting atomic notes from reviews, choosing review location preference

All prep notes created by this skill must include created-by: ai. Review notes are created-by: human. Extracted insights are created-by: ai-assisted. Never modify the body of the human’s review — only update frontmatter properties.

Integration with heartbeat

The heartbeat checks review cadences during STEP 8 — Skill opportunities (or a dedicated review step). Logic:

if today is Sunday (or >7 days since last weekly review):
    if fragments exist from the past week:
        suggest: "You have N fragments this week. Ready for a weekly review?"
        if human absent (unattended heartbeat): create prep note, leave board card

if today is 1st of month (or >30 days since last monthly review):
    if weekly reviews exist from the past month:
        suggest monthly review

if today is January (or >365 days since last yearly review):
    if monthly reviews exist from the past year:
        suggest yearly review

The heartbeat never writes the review — it creates the prep note and either prompts the human or leaves a board card (“Weekly review ready — see inbox/reviews/...“).

Integration with introspect

Fractal review and introspect are complementary practices:

During introspect’s LAYER 4 (Meta), reference the human’s recent reviews — they’re the richest signal for whether the meta layer accurately reflects the human’s actual thinking. A gap between what reviews say and what meta/ claims is a calibration opportunity.

Conventions

• The AI prepares; the human reflects. Never write the review for them.
• Prep notes are disposable — they live in inbox/reviews/ and can be cleaned up after the review is done.
• Review notes are sacred human documents (created-by: human). Never modify their body.
• If a cadence has no material (no fragments for daily, no weekly reviews for monthly), skip it. Don’t create empty reviews.
• The 40 questions template for yearly reviews comes from Kepano: https://stephango.com/40-questions
• Random revisit is part of the review process — always include 3–5 older notes that connect to current themes.
• Don’t over-prepare. The prep note should take <30 seconds to scan. If it’s too long, the human won’t read it.