triage

name	triage
description	Classify files in Clippings/ and route them to research/ or interview-prep/ subdirectories. Use at vault root when the user says "triage", "process clippings", or has accumulated unsorted clippings to process.

Triage

Parallel classification of files in Clippings/, routing each to research/raw/<sub>/ or interview-prep/raw/<sub>/.

Pattern: dispatch subagents → propose batch → user confirms → main agent executes.

User-facing output (table contents, summary report) defaults to the language the user writes in — mirror the user's input language. To pin a fixed language instead, tell your LLM at the start of the session, or edit this line.

Steps

1. List items

find Clippings -mindepth 1 -maxdepth 1 \( -type f -o -type d \)

Items can be either single files (most common) OR directories. Two common directory shapes:

• a multi-chapter book repo (honkit / mdbook source tree) → routes to research/raw/books/
• a cloned source-code repo kept verbatim (has package.json / src/ / build config, not just prose .md) → routes to research/raw/repos/ — distinct from code-notes/, which holds notes about a repo, not the repo itself

Treat each top-level item — file or directory — as one classification unit. Do not descend into directories to dispatch per-file classifiers.

If empty → report Clippings/ is empty, nothing to triage and stop.

2. Dispatch subagents in parallel

For each file, in a single message, issue multiple Agent tool calls with subagent_type: "triage-classifier" (loaded from agents/triage-classifier.md, or .claude/agents/triage-classifier.md if you're using Claude Code with the symlink layout). Each call passes the file's absolute path as input.

Each subagent returns a JSON object:

{
  "file": "<absolute path>",
  "action": "keep" | "delete",
  "destination": "<absolute path to target subdir>" | null,
  "reason": "<one English sentence>",
  "confidence": "high" | "medium" | "low"
}

Parallelism is critical. Do not call subagents serially — issue all Agent tool uses in one assistant turn.

3. Collect

Gather all returned JSON proposals. Split into two groups:

• Batch proposals — confidence of high or medium
• Needs human judgment — confidence of low, OR any subagent that flagged ambiguity in its reason

4. Present

Output a single table to the user:

## Triage Proposal

| File | Action | Destination | Reason |
|---|---|---|---|
| foo.md | keep | research/raw/blogs/ | LLM agent architecture analysis |
| bar.md | delete | -- | Shallow content, duplicates existing material |
| baz.md | keep | interview-prep/raw/system-design/ | "Design X" problem genre |

## ⚠ Needs your judgment

| File | Why uncertain | Candidates |
|---|---|---|
| qux.md | Looks like both research and interview-prep | research/raw/blogs/ or interview-prep/raw/companies/ |

Reply `go` to confirm, or edit any row first.

5. Wait for explicit confirmation

Do NOT execute any mv or rm before the user confirms.

Accepted confirmation signals: go, ok, confirm, execute. Ambiguous replies (looks fine, sure) → ask again for an explicit signal.

If the user edits any row → apply the edit, re-render the table, wait for confirmation again.

6. Execute in batch

After confirmation:

• mv all keepers in parallel (multiple mv in one Bash call, or multiple parallel Bash calls)
• rm all deletes in parallel

6.5 Verify execution

After the batch, confirm filesystem matches intent. Tool absence of error ≠ success — always check.

For each kept file:

test -e "<destination>"   # target should exist
test ! -e "<source>"      # source should be gone

For each deleted file:

test ! -e "<source>"      # source should be gone

Batch into one Bash call by chaining checks (echo any failures). On any failure:

• Auto-retry the failed mv/rm ONCE. Re-run the original command, then re-check.
• If retry succeeds → continue to step 7, but record which ops needed retry (signal something fragile in the source/destination — e.g., permission, symlink, name collision).
• If retry also fails → STOP. Do not proceed to logs. Report which file(s), the original command, and the error to the user. Ask whether to debug / abort / proceed-with-only-the-successful-ones.

Never proceed to step 7 without all verify checks passing (or explicit user override).

7. Update logs

For each kept file, append one line to the receiving subdomain's log.md (NOT the root):

## [YYYY-MM-DD] triage | <filename> → <destination relative to vault root>

Use today's date in YYYY-MM-DD format.

8. Report

Concise summary:

• Moved N, deleted M
• Whether Clippings/ is now empty (this is a healthy state — say so)
• Suggest next step: user can cd research/ to run ingest

Edge cases

• Cloned source-code repo: route the whole directory to research/raw/repos/ (verbatim code). Do not split it into per-file notes — that's the ingest skill's job, and the synthesized notes land in research/raw/code-notes/ / the wiki, not here.
• Neither research nor interview-prep fits: subagent returns low confidence. Surface in the ⚠ section. User decides (could go to a different subdomain like life/, stay in Clippings for later, or be deleted).
• Binary / unreadable files: subagent returns confidence: low with reason: "unreadable". Surface in ⚠ section.
• Filename mismatches content (e.g., a PDF named .md): subagent flags in reason, sets confidence low.
• Obvious media attachments (image/audio/video standalone in Clippings): probably an attachment that got separated from a parent doc — flag for user.

Do not

• Move or delete any file before user confirmation.
• Mix low-confidence items into the batch proposals — they belong in the ⚠ section.
• Read Clippings files yourself — let subagents do it. Keep main context clean.
• Write any wiki content — triage only routes; wiki updates are for the ingest skill.