Verified against ChatGPT · 2026-08-08
Turn a dumped-in inbox into ranked replies sorted by what decision each one actually needs
Pastes a batch of unanswered emails and gets each one triaged by decision type, with a draft reply for anything that doesn't need your personal judgment call.
The prompt
Ready to copy — highlighted parts are example details you can swap.
You are triaging a backlog of unanswered emails for me — not summarizing them, but sorting each one by what kind of decision it actually requires and drafting a reply for anything that doesn't need my personal judgment. EMAIL BACKLOG 5 emails: a vendor asking to reschedule a call, a client asking if a discount is still valid, a colleague asking for a doc link, a newsletter reply-all thread, and a partner proposing a new contract term. MY ROLE AND AUTHORITY I'm the ops manager — I can reschedule calls and answer process questions, but I can't approve discounts or new contract terms. STANDING DEFAULTS FOR ROUTINE REQUESTS Standard reschedule window is any weekday 10am-4pm; doc links live in the shared drive under /ops/current; newsletter threads get no reply. THINGS THAT MUST NEVER BE AUTO-REPLIED Anything mentioning a refund, legal threat, or a price different from the published rate card. For each email, first classify it into exactly one bucket: (1) routine — answerable from my standing defaults with no judgment call, (2) needs-my-call — the answer depends on a preference or trade-off only I can make, (3) needs-someone-else — I'm not the right person and it should be forwarded, or (4) no-reply-needed — informational only. Do not default everything into needs-my-call just to be safe; that defeats the point of triage. For bucket 1, write the full reply ready to send, using my standing defaults, and say which default you used. For bucket 2, do not draft a reply — instead write the one specific question I need to answer before a reply can exist, phrased so I can respond in a single word or sentence. For bucket 3, name who it should go to and a one-line forwarding note. For bucket 4, just confirm no action is needed and why. Never invent a commitment, price, date, or promise on my behalf that isn't backed by something explicitly in my standing defaults — if a routine-looking email actually requires a judgment call hiding inside it, reclassify it to bucket 2 rather than guessing. WHAT NOT TO DO Do not summarize the emails back to me in prose before triaging — go straight to the classification. Do not batch multiple unrelated emails into one combined reply even if they're from the same sender. OUTPUT FORMAT A table: Email (subject/sender), Bucket, Action (draft reply text, the one question, forward-to name, or "no action"). End with a one-line count of how many landed in each bucket.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
The core failure mode of asking an LLM to "help with my inbox" is that it treats every email with the same generic helpfulness, which either produces a wall of drafted replies you now have to individually vet for made-up commitments, or a wall of summaries that just re-describes the problem back to you without reducing your workload. Forcing an explicit four-way classification before any drafting happens changes the task from open-ended assistance into a bounded sorting problem, which GPT-5.1 handles far more reliably than open-ended judgment — it's much better at applying a stated rule (does this match a standing default, yes or no) than at silently inferring when it should defer to you. The instruction to surface a single answerable question for bucket 2, rather than drafting a placeholder reply, matters because a half-drafted reply with a guessed answer is more dangerous than no draft at all — it looks finished, so it's more likely to get sent without a real check, whereas a bare question forces you to actually supply the missing judgment before anything goes out. Explicitly naming a never-auto-reply category closes a specific failure mode: the model treating a phrase like "is the discount still valid" as routine because the surrounding email reads politely, when in fact any dollar-figure deviation from a rate card is exactly the kind of thing that shouldn't be answered on autopilot. The reclassification instruction — if a routine-looking email hides a judgment call, move it to bucket 2 rather than guessing — exists because the model's default failure under time pressure is to force borderline cases into whichever bucket lets it produce more finished-looking output, and this explicitly removes that incentive.
What you get back
| Email | Bucket | Action | |---|---|---| | Vendor reschedule request | Routine | Draft: "Happy to move our call — I'm free any weekday 10am-4pm, let me know what works." (used reschedule-window default) | | Client discount question | Needs-my-call | Question: is the 15% discount still valid past the original quote date, yes or no? | | Colleague doc link | Routine | Draft: link to /ops/current shared drive folder |
Bucket counts: 2 routine, 1 needs-my-call, 1 needs-someone-else, 1 no-reply-needed.
Verified against
ChatGPT GPT-5.1 · 2026-08-08
Changelog
- 2026-08-08 — Initial publish, verified against ChatGPT GPT-5.1.
Need this built into your business?
If a prompt isn't enough — custom software, built and maintained for you — that's Scult's day job.
EXPLORE CUSTOM SOFTWARE
