# Triage Workflow

Batch-process job alert emails into a living Job Triage document.

**Trigger:** "evaluate the next N mails", "triage jobs", "batch analyze", "check job alerts"

---

## Prerequisites

1. **Switch SL user**: Call `sl_switch_user(name: "matthias")` before any SL API calls
2. **Load preferences**: Call `sl_get_preferences()` for evaluation criteria
3. **Determine triage file**: `Notes/YYYY/MM/Job Triage.md` (current month). Create if missing.

---

## Step 1: Find Job Alert Emails

Search Gmail for recent LinkedIn job alerts (or other sources as specified):

```
gmail search: from:linkedin.com newer_than:Nd in:inbox subject:job
```

- Use `search_gmail_messages` to get message IDs
- Default: last 7 days unless user specifies otherwise
- Identify which emails are job alert digests (vs. connection requests, notifications)

**IMPORTANT:** Do NOT rely on subjects or metadata snippets. Proceed to Step 2 immediately for full body reads.

---

## Step 2: Read Full Email Bodies and Extract ALL Job IDs

### MANDATORY: Full content read for every email

For EVERY job alert email identified in Step 1, call `get_gmail_message_content(message_id)` to read the COMPLETE email body. Never skip this step. Never rely on metadata, snippet, or subject line alone.

### Extract ALL LinkedIn job URLs

Parse the full body text for every URL matching this pattern:

```
linkedin.com/jobs/view/XXXXXXXXXX
```

Extract every single job ID found. Do not stop at the first match. Do not skip URLs that look similar to others already found.

### Per-email manifest (REQUIRED)

Before moving on, record a manifest entry for every email processed:

```
Email: <subject> (<date>)
Message ID: <id>
Jobs extracted: <count>
Job IDs: <comma-separated list of LinkedIn job IDs>
```

Log this to the session notes or a scratch variable. You will use it for the reconciliation report in Step 8.

### Quality gate

- If an email yields 0 extracted job IDs, flag it explicitly as a PARSING FAILURE - do not silently skip it
- LinkedIn job alert emails typically contain 5-15 job listings; fewer than 3 may indicate extraction failed
- Re-read the raw body if extraction seems low

### Build the master job list

After processing ALL emails, deduplicate by job ID. The master list is your working set for Steps 3-6.

---

## Step 3: Dedup Against Existing Triage

Before evaluating any job, check each job ID (or company + title) against the current month's Job Triage file:

- If the job is already in the Active Pipeline or Dismissed table, skip it
- Note how many were skipped as duplicates for the reconciliation report

This prevents re-evaluating jobs already processed in a previous session.

---

## Step 4: Quick Triage (Title/Location Filter)

Before scraping full JDs, do a fast pass to eliminate obvious mismatches:

**Auto-reject (score 1/5, Triage: No):**
- Wrong country (US, UK, etc. - unless remote with CH contract)
- Pure developer/engineer roles (frontend, backend, full-stack, embedded, ML engineer)
- Academic positions (postdoc, PhD, graduate assistant)
- Apprenticeships, internships, stagiaire
- Hardware/electronics/mechanical engineering
- Roles clearly below senior level
- Salary explicitly below CHF 150k

**Keep for full evaluation (score TBD):**
- IT leadership, management, director, head of
- Consulting, advisory, business analyst (senior)
- Program/project management (IT/digital)
- Product management (tech companies)
- Architecture, strategy, transformation
- Suisse romande or remote-CH

---

## Step 5: Scrape Full JDs (Parallel)

For jobs that pass the quick triage (scored >= 2/5):

**Scraping priority order:**
1. **Scribe** `extract_content` - try first (fast, no browser needed)
2. **Playwright** - if Scribe returns 403/empty (required for Glassdoor, some LinkedIn)
   - `browser_navigate` to URL
   - `browser_evaluate` with appropriate JS selector:
     - Glassdoor: `document.querySelector('[class*="JobDetails_jobDescription"]').innerText.substring(0,5000)`
     - LinkedIn: `document.querySelector('main').innerText.substring(0,5000)`
3. **sl_scrape_job** - for supported job sites (job-room.ch, LinkedIn)

Launch parallel agents for scraping when possible (one per URL).

---

## Step 6: Evaluate Each Job

For each scraped JD, evaluate against the user's profile:

**Scoring criteria (from preferences):**
- IT leadership / consulting alignment (weight: high)
- Seniority level (must be senior/director/head)
- Domain fit (IT, digital transformation, ERP, data, cloud)
- Location (Suisse romande preferred, hybrid/remote OK)
- Salary (floor CHF 150k)
- Company type (consulting engagement vs permanent - note which)
- Language requirements (FR/EN/DE - all OK)

**Score scale:**
- **5/5**: Perfect match - IT leadership, right location, right level
- **4/5**: Strong match with minor gap (e.g., slightly different domain)
- **3.5/5**: Good match, worth pursuing (consulting roles, senior scope)
- **3/5**: Possible match, needs closer look
- **2/5**: Partial match, significant gaps
- **1/5**: Clear mismatch (wrong domain, level, location, or role type)

**BatchAnalyze threshold:** Only create jobs in SeriousLetter for scores >= 3/5.

---

## Step 7: Update Job Triage File

### File Structure

The Job Triage file (`Notes/YYYY/MM/Job Triage.md`) has this structure:

```markdown
# Job Triage

Living document. Jobs flow through here: new batch -> triage -> pursue or dismiss.
- **Pursue**: detail section gets fleshed out, moves to own 0XXX folder when applied
- **Dismiss**: row removed (or moved to Dismissed section at bottom)
- **Goal**: always shows what's active right now

---

## Active Pipeline

| #   | Company                               | Title                          | Location  | Score | Triage | Notes              |
| --- | ------------------------------------- | ------------------------------ | --------- | ----- | ------ | ------------------ |
| 1   | [[#Company Name\|Company Name]]       | Job Title Here                 | City      | 3/5   | No     | Brief reason       |

---

## Completed

| # | Company | Title | Outcome | Folder |
|---|---------|-------|---------|--------|

---

## Dismissed

| # | Company | Title | Reason |
|---|---------|-------|--------|

---

## Detail Sections

---

### Company Name

**Location:** City | **Salary:** Not listed | **Score:** X/5
**Source:** [Glassdoor](https://...)

**Why this fits:**
Explanation of alignment with profile...

**Gaps/concerns:**
What doesn't match or needs investigation...

**Next step:** Scrape full JD / Create in SL / Skip
```

### Table Row Format

Each job gets a numbered row in the Active Pipeline table:
- **#**: Sequential within the file (1, 2, 3...)
- **Company**: Obsidian wiki-link to detail section: `[[#Company Name\|Company Name]]`
  - If company appears multiple times, use a unique anchor: `[[#Company-suffix\|Company]]`
- **Title**: Job title (keep concise, max ~50 chars)
- **Location**: City or region
- **Score**: X/5 rating
- **Triage**: "Yes" (pursue), "No" (skip), or empty (not yet evaluated)
- **Notes**: 2-4 word reason for the score/decision

### Detail Section Format

Each job gets a detail section below `## Detail Sections`:

```markdown
---

### Company Name

**Location:** City | **Salary:** CHF XXXk or "Not listed" | **Score:** X/5
**Source:** [Site Name](URL)

**Why this fits:**
2-3 sentences on alignment with IT leadership/consulting profile.

**Gaps/concerns:**
1-2 sentences on mismatches or unknowns.

**Next step:** What to do next (Scrape full JD / Create in SeriousLetter / Skip)
```

For jobs scoring >= 3/5 with full JD scraped, add more detail:

```markdown
---

### Company Name

**Location:** City (hybrid/remote) | **Salary:** CHF XXXk | **Score:** X/5 | **Triage:** Yes
**Source:** [Site Name](URL)
**Type:** Permanent / Consulting engagement

**Full JD summary:** 2-3 sentence summary of the role, responsibilities, requirements.

**Why this fits:**
- Bullet point 1
- Bullet point 2
- Bullet point 3

**Gaps/concerns:**
- Bullet point 1
- Bullet point 2

**Next step:** Create in SeriousLetter, write cover letter, apply
```

### Creating a New Triage File

If `Notes/YYYY/MM/Job Triage.md` doesn't exist for the current month:

1. Create the file with the full template (header, empty Active Pipeline table, empty Completed/Dismissed, Detail Sections header)
2. Start numbering from 1

### Appending to Existing Triage File

If the file already exists:

1. Read the current file
2. Find the highest # in the Active Pipeline table
3. Append new rows starting from max + 1
4. Append new detail sections at the bottom (before the last `---` or at end)
5. Do NOT renumber existing rows
6. Do NOT modify existing detail sections

---

## Step 8: Archive Processed Emails

Archive each Gmail job alert email ONLY AFTER all jobs from that email have been extracted and recorded in the manifest. Do not archive early.

- Remove INBOX label
- Add Jobs/* label (e.g., Jobs/01 - Pipeline)

---

## Step 9: Summary and Reconciliation Report

Present a summary table to the user:

```
| # | Company | Position | Fit | Action |
|---|---------|----------|-----|--------|
| 1 | Acme    | CTO      | 4/5 | Create in SL |
| 2 | BigCo   | Dev Lead | 2/5 | Skip   |
```

Then present the per-email manifest so the user can spot any gaps:

```
Email manifest:
- "Your job alert for CTO roles" (2026-03-09): 8 jobs extracted
- "Weekly digest: IT leadership" (2026-03-08): 12 jobs extracted
- "New matches for you" (2026-03-07): 0 jobs extracted *** PARSING FAILURE - review manually ***
```

Reconciliation totals:
- Total emails processed: N
- Total job IDs extracted (before dedup): N
- Duplicate / already-known: N
- New jobs evaluated this session: N
- Jobs worth pursuing (>= 3/5): N
- Jobs skipped (< 3/5): N
- Scraping failures: N
- Parsing failures (0 jobs from non-empty email): N (flag each one)

**Extraction rate check:** If (total jobs extracted / emails processed) < 3, warn the user that extraction may have failed on one or more emails.

---

## Important Rules

- **NEVER use em-dashes** in any generated content - use hyphens with spaces
- **ALWAYS read full email body**: Use `get_gmail_message_content`, never rely on metadata/snippet alone
- **ALWAYS extract ALL job IDs**: Parse every `linkedin.com/jobs/view/XXXXXXXXXX` URL from the full body
- **ALWAYS produce per-email manifest**: Record subject, date, message ID, and count of jobs extracted per email
- **ALWAYS dedup first**: Cross-check against existing triage file before evaluating
- **ARCHIVE AFTER extraction**: Never label/archive an email before all its jobs are recorded
- **Obsidian links**: Always use `[[#Heading\|Display]]` format in tables (standard markdown anchors don't work in Obsidian tables)
- **Parallel scraping**: Launch multiple Playwright agents when scraping 3+ URLs
- **Gmail archiving**: Always remove INBOX label when adding Jobs/* labels
- **File is append-only**: Never remove or modify existing entries unless user explicitly asks

---

**Last Updated:** 2026-03-09