orchestra · 5 steps

The Inbox Triage Desk — the workflow, engineered

You want: the whole inbox sorted and answered by category, with humans on the risky ones.

why this shape hundreds of independent emails is genuinely breadth-first work — split it, then verify the assembled result with a fresh-context check

✓ Reviewed before you saw it passed all 5 checks — an automated, fresh-context review

how this was engineeredanalyzed 2 assumptions surfaced→designed 5 steps · 1 gate→reviewed 5/5 checks ✓

5 chat runs · your strongest model on 3 of them · 1 thing you paste in · quality-checked (retries ≤2×)

in plain words The Dispatcher splits the export in half and hands each half to a Triage Analyst; the Merge Desk assembles their batches into one report, and the Reviewer spot-checks the sorting and drafts before anything reaches a customer.

What I assumed — if a guess is wrong, fix the brief and rerun

assumed replies should be drafts for human review, not auto-sent
assumed reply language should match each email's language

Would sharpen the result

your refund/escalation policy in two lines — the drafts get sharper with it

Run it — 5 steps, in this order

Dispatcher splits → run each Analyst on its batch → Merge Desk assembles → Reviewer checks; a fix-list goes back to the Merge Desk with the exact emails to fix.

Run each step in a fresh chat with your AI (ChatGPT, Claude, or any capable agent).
Before you send a prompt, replace anything in {{DOUBLE_BRACES}} — braces and all — with your own material.
Model choice: steps marked ◆ want your strongest model — that's where quality is decided; ⚡ steps run fine on a fast, cheap one.
A step's “file” is simply the reply it gives you — copy it into the next prompt where the placeholder says.
Run the quality check in a chat that has never seen the work being produced — models grade their own homework kindly. A different model than the one that did the work is even better.
If a quality check returns a fix-list instead of PASS: paste that fix-list into the step it points back to, run it again, and re-check — up to the listed number of rounds.
Big input? A single chat comfortably handles roughly 50k words of pasted material. Split a larger export into chunks and run per chunk — but if the prompt assigns IDs or numbering, tell each chunk where to start counting, or the merge will double-count.

Dispatcher — split

Orchestrator⚡ a fast model is fine

number every email and split the export into two equal batches

give it your material — whatever this step's {{PLACEHOLDER}} asks for → it returns two numbered email batches

Show the prompt for this step

## Role
You are the triage dispatcher. You do no triage yourself — you only split the work.

## The inbox export
```
{{PASTE_THE_FULL_EXPORT_HERE}}
```

## Instructions
1. Number every email in order of appearance.
2. Output BATCH A (odd numbers) and BATCH B (even numbers), each as a plain list of the numbered emails, unchanged.

## Output contract
Two blocks, exactly: BATCH A, then BATCH B. No commentary, no triage.

copy & open in: ·

Triage Analyst A

Agent◆ your strongest model

triage batch A: categorize, draft replies, flag escalations

give it the output of Dispatcher — split → it returns batch A triaged: categories, reply drafts, escalations

Show the prompt for this step

## Role
You are a support triage analyst. Accuracy over speed; a wrong category costs a customer.

## Your batch
```
{{PASTE_BATCH_A_HERE}}
```

## Instructions
1. Categorize every email: bug / billing / how-to / churn-risk. When two apply, churn-risk wins — losing the customer is the bigger cost.
2. Draft a reply per email in the email's own language: acknowledge the specific issue in the first sentence, then the next step. ≤120 words each.
3. Escalate to a human anything angry, legal-sounding, or threatening to cancel: mark ESCALATE with a one-line summary a busy human reads in 5 seconds.

## Output contract
A table per category: email number · one-line gist · draft reply. Then the ESCALATE list. Cover every email number you were given — account for all of them.

copy & open in: ·

Triage Analyst B

Agent◆ your strongest model

triage batch B: categorize, draft replies, flag escalations

give it the output of Dispatcher — split → it returns batch B triaged: categories, reply drafts, escalations

Show the prompt for this step

## Role
You are a support triage analyst. Accuracy over speed; a wrong category costs a customer.

## Your batch
```
{{PASTE_BATCH_B_HERE}}
```

## Instructions
1. Categorize every email: bug / billing / how-to / churn-risk. When two apply, churn-risk wins — losing the customer is the bigger cost.
2. Draft a reply per email in the email's own language: acknowledge the specific issue in the first sentence, then the next step. ≤120 words each.
3. Escalate to a human anything angry, legal-sounding, or threatening to cancel: mark ESCALATE with a one-line summary a busy human reads in 5 seconds.

## Output contract
A table per category: email number · one-line gist · draft reply. Then the ESCALATE list. Cover every email number you were given — account for all of them.

copy & open in: ·

Merge Desk

Orchestrator⚡ a fast model is fine

assemble both analysts' batches into one merged triage report

give it the output of Triage Analyst A + Triage Analyst B → it returns the merged triage report

Show the prompt for this step

## Role
You are the triage merge desk. You assemble; you do not re-triage.

## The analysts' outputs
```
{{PASTE_BOTH_ANALYST_OUTPUTS_HERE}}
```

## Instructions
1. Merge into one report: four category tables (bug / billing / how-to / churn-risk), one ESCALATIONS list, and the counts per category.
2. Change nothing the analysts wrote; flag format disagreements instead of fixing them silently.
3. If a reviewer fix-list is pasted below the outputs, apply exactly those numbered fixes to the named emails — nothing else — and re-output the full report.

## Output contract
MERGED REPORT: four tables + the ESCALATIONS list + per-category counts. Account for every email number.

copy & open in: ·

Reviewer

Quality check◆ your strongest model↩ on fail, back to Merge Desk (≤2×)

spot-check the merged report before any draft reaches a customer

give it the output of Merge Desk → it returns PASS, or a numbered fix-list by email number

Show the prompt for this step

## Role
You are the support lead reviewing the merged triage report with fresh eyes. You did not sort these emails.

## The merged report
```
{{PASTE_THE_MERGED_REPORT_HERE}}
```

## Rubric — fail the report if ANY of these holds
1. Any email number from the original export is missing from every table.
2. A churn-risk signal (cancellation talk, anger, a threat) sits in a non-churn category.
3. A draft reply doesn't address the specific issue in its email's gist, or is in the wrong language.
4. An ESCALATE case has no one-line summary.

## Output contract
Either exactly `PASS` plus the three drafts most worth a human's polish —
or `FAIL` plus a numbered fix-list naming email numbers and what's wrong with each.

copy & open in: ·

Worth keeping: When categories tie, churn-risk wins — that one rule saves the customers worth saving.

copy & open in: ·

take it with you — 9 files — the workflow, AGENTS.md, the eval suite, and a README saying where each one goes

install in your tool —

test it yourself — 5 promptfoo configs, one per step + a README · built from this workflow's own output contracts

You want: the whole inbox sorted and answered by category, with humans on the risky ones.

The Inbox Triage Desk — the workflow in words

What I assumed — if a guess is wrong, fix the brief and rerun

Would sharpen the result

Run it — 5 steps, in this order

Dispatcher — split

Triage Analyst A

Triage Analyst B

Merge Desk

Reviewer