r007b34r inbox-triage
Constraints
- NEVER scan whole vault; scope =
Inbox/or user-specified path only - NEVER write back without explicit user confirmation
- NEVER produce synthesis or thesis; output is routing decisions only
- NEVER deep-read all items; deep-read only when routing is genuinely uncertain
- MAX 20 items per batch; split and report remainder if more
Trigger Boundary
Use when: triage inbox, sort captures, route raw notes, process fleeting notes, decide what to promote/archive.
Do NOT use for:
- Cross-note relationship analysis →
connection-review - Week-level meaning extraction →
weekly-synthesis - Stable note creation →
note-promotion - Vault health diagnosis →
vault-health-feedback
Boundaries
- Owns: inbox routing, capture classification, shallow-read decisions
- Delegates to:
connection-review(items needing links),note-promotion(items marked promote) - Never absorbs: synthesis, health diagnosis, context maintenance, relationship analysis
Continuations
| Condition | Next skill |
|---|---|
| Promoted items with link potential > 2 | connection-review |
| Item marked promote-to-note, user confirms | note-promotion |
| Inbox path not found or empty | vault-health-feedback |
| Ambiguous items > 60% of batch | Escalate to human |
Companion Skill
All vault operations follow obsidian-mcp. Forbidden: obsidian_patch_note, obsidian_append_to_note. Verify every write through readback.
Procedure
Phase 1: Discover
obsidian_list_notes path:"Inbox/"→ candidate list- Count items; if >20, process first 20 and report remainder
Phase 2: Shallow Read
- For each:
obsidian_get_note format:"document-map"→ structure - Short notes (<500 words): read as
format:"content" - Flag genuinely uncertain items → mark
needs-deep-read
Phase 3: Classify
When classifying, read references/examples.md for decision calibration.
| Decision | Use when |
|---|---|
| keep-in-inbox | Too raw, still incubating |
| promote-to-note | Stable, reusable, independent |
| promote-to-idea | Personal observation worth preserving |
| move-to-project | Mainly useful inside active project |
| archive-or-ignore | Low-density, redundant, stale |
| needs-deep-read | Shallow evidence insufficient, potential value high |
Dimensions: reusability, current relevance, cognitive density, connection potential.
Phase 4: Report
When formatting output, read references/templates.md for the exact output format.
Output batch summary per contract. Wait for user confirmation before any write-back.
- Empty inbox: report empty, suggest checking path config
- All items ambiguous: escalate to user, ask for 2-3 example decisions to calibrate
- Mixed languages: classify by content quality, not language
- Broken frontmatter: flag but still classify content
- Oversized inbox (50+): process first 20, report "X items remaining, continue?"
- Items with no content (title only): classify as needs-deep-read or archive
Gotchas
Gotcha 1: Triage becomes mini-synthesis
What happens: Agent writes multi-paragraph analysis for each item Why it's wrong: Triage is routing, not content production; wastes tokens and crosses into weekly-synthesis territory Correct approach: One sentence reason + one decision per item
Gotcha 2: Over-promotion
What happens: Agent marks most items as promote-to-note Why it's wrong: Promotes unstable fragments, creates hollow vault growth Correct approach: Promote only when stable + reusable + independent; default to keep-in-inbox when uncertain
Gotcha 3: Full-reading everything
What happens: Agent reads every note in full before classifying Why it's wrong: Wastes token budget; most routing decisions need only document-map Correct approach: Document-map first; deep-read only genuinely uncertain items
Validators (CI only, not runtime)
validators/pre-check.sh: Confirms inbox path accessible and not emptyvalidators/post-check.sh: Verifies output contains no synthesis markers
Exit Criteria
- Every candidate has exactly one routing decision or needs-deep-read
- Triage has not turned into synthesis
- Batch size ≤20 respected
- Any write-back passed obsidian-mcp readback verification