triage-inbox

name	triage-inbox
description	Classify a recruiting email (or a batch) into one of the 10 canonical email classes, extract structured fields (company / role / dates / links / asks / deadline / sentiment), and suggest the next action. Writes inbox/triage-<YYYY-MM-DD>.json — consumed by infer-status and draft-reply.
when_to_use	Use when the user pastes one recruiter email OR JobClaw passes a batch from the AgentMail inbox. Front of the recruiter loop.
user-invocable	true
allowed-tools	Read, Write

triage-inbox

Tell the user (or JobClaw) what each email is and what to do about it. Hard rules: ../_shared/RULES.md. All enums, the decision table, the extraction-fields shape, and the suggested-action vocabulary live in ../../knowledge/status/taxonomy.md — read it first; don't re-derive.

Inputs

Single mode

• A pasted email: from, subject, body, optional thread (prior messages), optional date.

Batch mode

• --in inbox/incoming.json:

  { "emails": [ { "email_id": "...", "from": "...", "subject": "...", "body": "...", "date": "YYYY-MM-DD", "thread": ["..."] } ] }
  

If the input has only a body (no from/subject), warn but proceed — class confidence will be low unless body is decisive.

Method

0. Read the taxonomy. ../../knowledge/status/taxonomy.md — the 10 classes, the extraction-fields contract, the suggested-action vocab.

1. For each email, classify into exactly one class. Mutually exclusive; pick the most specific match (e.g., a "we received your application" auto-message is info, not screen; a take-home request is assessment, not interview-invite). Use the body + subject + sender domain together — sender alone is unreliable.

2. Extract fields per the taxonomy's Extraction fields block (only fields the email genuinely provides — never invent):

   • company, role, job_id
   • dates[] — proposed slots / deadlines / start dates (ISO + time + tz + label)
   • links[] — { url, kind ∈ schedule|assessment|portal|other }
   • asks[] — verbatim recruiter asks ("send updated CV", "share notice period")
   • deadline — single hard cutoff if explicit
   • sentiment ∈ {neutral, positive, negative}
   • confidence ∈ {high, medium, low} — your own confidence in the classification

3. Suggest one action from the taxonomy's vocab (reply-schedule | reply-accept | reply-decline | reply-ask | reply-info | acknowledge | ignore | escalate). Map per the decision table; always suggest escalate for class other or any classification with confidence: low.

4. Apply guardrails.

   • offer emails → suggest reply-ask (acknowledge + buy time + recommend /coach-negotiation). Never suggest reply-accept directly.
   • assessment / interview-invite emails with a deadline within 48h → bump the email's priority field to urgent in the output (extension on the schema; downstream tooling can sort by it).
   • If the body looks like phishing (suspicious domain, prize wording, urgency + a link) → class spam regardless of subject.

Output

inbox/triage-<YYYY-MM-DD>.json

{
  "as_of": "YYYY-MM-DD",
  "triaged": [
    {
      "email_id": "...",
      "from": "...",
      "subject": "...",
      "class": "rejection | screen | interview-invite | assessment | offer | recruiter-outbound | info | request-info | spam | other",
      "extraction": {
        "company": "...", "role": "...", "job_id": "...",
        "dates": [...], "links": [...], "asks": [...],
        "deadline": "YYYY-MM-DD | null",
        "sentiment": "neutral | positive | negative",
        "confidence": "high | medium | low"
      },
      "suggested_action": "...",
      "priority": "normal | urgent",
      "notes": "string | null"
    }
  ]
}

Stdout

A concise table per email: id · class · company/role · action · priority · confidence. Highlight urgent rows and any escalate rows for the user to review.

Anti-patterns

• Treating "we received your application" as screen (it's info).
• Collapsing screen and interview-invite — they trigger different downstream actions.
• Inventing job_id, dates, or company not in the source text.
• Setting confidence: high on ambiguous emails just to clear the queue. Be honest — low is more useful.
• Suggesting reply-accept for an offer (must be reply-ask + recommend /coach-negotiation).
• Emoji.

Rubric

• Class is one of the 10 taxonomy values, exactly.
• Every extraction.* field is sourced from the email (or null/[]).
• Suggested action is from the taxonomy vocab and consistent with the decision table.
• No offer is paired with reply-accept.
• confidence: low rows are flagged with suggested_action: escalate.
• Urgent items (assessment / interview-invite with ≤48h deadline) marked priority: urgent.

Next steps

Triaged. Next: /infer-status to update the funnel, then /draft-reply for the suggested action.